Research article

Density investigation and implications for exploring iron-ore deposits using gravity method in the Hamersley Province, Western Australia

  • Received: 14 November 2022 Revised: 22 November 2022 Accepted: 28 November 2022 Published: 05 December 2022
  • The Hamersley Province in the northwest of Western Australia contains extensive banded iron formations (BIFs) and large hematite-goethite deposits. Density information of rocks and ores in this region has been scarce. This study reports the results of a systematic density investigations based on more than eight hundred density datasets in the province. This study not only provides a better understanding of density distribution of the rocks and ores in the province, but also allows forward gravity modeling over the known iron-ore deposits to be conducted for exploring the usefulness and effectiveness of gravity surveys for detecting concealed iron-ore deposits in the region. This should have a significant impact on iron-ore mining in the province as the outcropped ores have been mined for over 40 years in the province and the future targets are likely the concealed deposits below the surface. The analysis shows a clear density contrast around 1.0 g/cm3 between the Brockman iron ores and the host BIFs, which should generate clear positive net gravity anomalies over buried large iron-ore deposits. However, porous goethite ores hosted in the Marra Mamba BIFs have an average density of about 2.8 g/cm3 due to porosity about 30–40% in the ores. A density contrast of −0.5 g/cm3 may exist between the goethite ores and BIFs, which would produce net negative gravity anomalies over the deposits. Since most goethite deposits are layered consistently with the host rocks and associated with broad folds, the net gravity anomaly of an orebody itself may generally have the similar shape to the corresponding BIF bedrock. This implies that gravity surveys may be able to detect paleochannels which host the goethite ores, rather than directly detecting the orebody.

    Citation: William Guo. Density investigation and implications for exploring iron-ore deposits using gravity method in the Hamersley Province, Western Australia[J]. AIMS Geosciences, 2023, 9(1): 34-48. doi: 10.3934/geosci.2023003

    Related Papers:

    [1] Hicham Saber, Tariq Alraqad, Rashid Abu-Dawwas . On graded s-prime submodules. AIMS Mathematics, 2021, 6(3): 2510-2524. doi: 10.3934/math.2021152
    [2] Malak Alnimer, Khaldoun Al-Zoubi, Mohammed Al-Dolat . On graded weakly Jgr-semiprime submodules. AIMS Mathematics, 2024, 9(5): 12315-12322. doi: 10.3934/math.2024602
    [3] Shatha Alghueiri, Khaldoun Al-Zoubi . On graded 2-absorbing Ie-prime submodules of graded modules over graded commutative rings. AIMS Mathematics, 2020, 5(6): 7624-7631. doi: 10.3934/math.2020487
    [4] Tariq Alraqad, Hicham Saber, Rashid Abu-Dawwas . Intersection graphs of graded ideals of graded rings. AIMS Mathematics, 2021, 6(10): 10355-10368. doi: 10.3934/math.2021600
    [5] Sami Ul Haq, Saeed Ullah Jan, Syed Inayat Ali Shah, Ilyas Khan, Jagdev Singh . Heat and mass transfer of fractional second grade fluid with slippage and ramped wall temperature using Caputo-Fabrizio fractional derivative approach. AIMS Mathematics, 2020, 5(4): 3056-3088. doi: 10.3934/math.2020198
    [6] Zehba Raizah, Arshad Khan, Saadat Hussain Awan, Anwar Saeed, Ahmed M. Galal, Wajaree Weera . Time-dependent fractional second-grade fluid flow through a channel influenced by unsteady motion of a bottom plate. AIMS Mathematics, 2023, 8(1): 423-446. doi: 10.3934/math.2023020
    [7] Yu-Hsien Liao . Integrating axiomatic and dynamic mechanisms under industrial management situations: game-theoretical analysis. AIMS Mathematics, 2025, 10(6): 14975-14995. doi: 10.3934/math.2025671
    [8] Mir Sajjad Hashemi, Hadi Rezazadeh, Hassan Almusawa, Hijaz Ahmad . A Lie group integrator to solve the hydromagnetic stagnation point flow of a second grade fluid over a stretching sheet. AIMS Mathematics, 2021, 6(12): 13392-13406. doi: 10.3934/math.2021775
    [9] Imran Siddique, Yasir Khan, Muhammad Nadeem, Jan Awrejcewicz, Muhammad Bilal . Significance of heat transfer for second-grade fuzzy hybrid nanofluid flow over a stretching/shrinking Riga wedge. AIMS Mathematics, 2023, 8(1): 295-316. doi: 10.3934/math.2023014
    [10] Takao Komatsu, Ram Krishna Pandey . On hypergeometric Cauchy numbers of higher grade. AIMS Mathematics, 2021, 6(7): 6630-6646. doi: 10.3934/math.2021390
  • The Hamersley Province in the northwest of Western Australia contains extensive banded iron formations (BIFs) and large hematite-goethite deposits. Density information of rocks and ores in this region has been scarce. This study reports the results of a systematic density investigations based on more than eight hundred density datasets in the province. This study not only provides a better understanding of density distribution of the rocks and ores in the province, but also allows forward gravity modeling over the known iron-ore deposits to be conducted for exploring the usefulness and effectiveness of gravity surveys for detecting concealed iron-ore deposits in the region. This should have a significant impact on iron-ore mining in the province as the outcropped ores have been mined for over 40 years in the province and the future targets are likely the concealed deposits below the surface. The analysis shows a clear density contrast around 1.0 g/cm3 between the Brockman iron ores and the host BIFs, which should generate clear positive net gravity anomalies over buried large iron-ore deposits. However, porous goethite ores hosted in the Marra Mamba BIFs have an average density of about 2.8 g/cm3 due to porosity about 30–40% in the ores. A density contrast of −0.5 g/cm3 may exist between the goethite ores and BIFs, which would produce net negative gravity anomalies over the deposits. Since most goethite deposits are layered consistently with the host rocks and associated with broad folds, the net gravity anomaly of an orebody itself may generally have the similar shape to the corresponding BIF bedrock. This implies that gravity surveys may be able to detect paleochannels which host the goethite ores, rather than directly detecting the orebody.



    The energy-management functions and timing have been greatly influenced by load and power prediction. An effective load forecasting and energy management system is essential to a fair and sustainable system. Many variables, including cost, load flow and power plant demand, depend on data collected by the load and power forecasting system [1,2]. Large energy storage and demand fluctuations also require a proactive balancing of load changes and energy management systems, as shown in [3,4]. A reliable load and power forecasting management system help reduce power and load forecast errors, enhance the robustness of power systems and mitigate changes in supply and demand [5,6].

    Conventional statistical approaches including linear regression analysis (LR) approach and autoregressive moving average (ARMA) methodology have been previously used in pregnancy prediction methods based on massive amounts of data [7,8]. However, traditional statistical techniques cannot adequately learn these nonlinear data or match pregnancy prediction accuracy criteria due to many intrinsic nonlinear properties of big data [9,10]. Modern artificial intelligence technology enables energy systems to analyze data and manage demand properly. Machine learning methods have recently been used to predict power consumption and load patterns. These approaches lead to better control of load and power resulting in lower energy use, longer resource life and lower processing costs [11,12,13].

    Currently, neural networks are being used more and more to tackle complex time-series prediction problems as a result of their recent success in challenges associated with computer vision and image processing. By comparing actual usage with forecasts from the forecasting system, accuracy can be improved as well as money and energy savings. Scheduling and planning of power systems can be damaged by inaccurate load forecasting which can lead to power shortages in the market and eventually increase the desire to expand existing Artificial Neural Network (ANN) designs [14,15]. By adding more hidden layers, which may handle nonlinearities in the input and output data, the extension is produced. The goal is to provide a Deep Neural Network (DNN) based power load prediction model for power usage prediction.

    Initially, many neural networks used backpropagation techniques for short-term load forecasting as proposed in [16,17,18]. Similarly, the processing of sequence data is done in many works using a recurrent neural network proposed in [19,20,21]. Based on climatological data, a local and Global RNN model was developed to address the issue of long-term wind speed and energy forecasting [22,23]. However, during error backpropagation, the gradient descent problem still exists due to the enormous depth of time in the traditional Recurrent Neural Network (RNN) architecture which cannot learn the long-term historical data. Comparatively, long short-term memory (LSTM), an improved form of RNN was introduced in [24] which uses gated mechanisms to integrate short-term and long-term memory.

    Latterly, Gated Recurrent Unit (GRU), based on an optimized LSTM combines the input gate and forget LSTM gate into a single update gate [25,26]. Still, the GRU model is not enough, because, the power system additionally contains high-dimensional data such as image data, spatiotemporal matrices and sequence data. These optimized models are not sufficient to handle high-dimensional data effectively. By contrast, a CNN is suitable for processing large amounts of high-dimensional data and has been successfully applied in many research works [27,28]. The CNN may record the characteristics of the local orientation and the non-variable features of the scale when there is a significant correlation between adjacent data points [29,30].

    Accordingly, there are four categories of power load projections: long-term, medium-term, short-term and very short-term [31,32,33,34]. The main focus should be on very short-term electrical load and power forecasting, which predicts future loads and may assist power system operators set up realistic production plans, keeping supply and demand in balance, and assuring safety while minimizing resource waste and electricity prices. The conventional statistical methodology was proven to be ineffective for learning nonlinearity due to the multiple intrinsic nonlinear features of large and diverse data sets. Therefore, to predict energy consumption, machine and deep learning approaches have recently been applied to solve issues related with large dimensionality and learning long-term dependencies. A novel load and energy demand predictor using a multi-directional GRU (MD-GRU) and CNN using different pooling schemes has been proposed in this work. To improve the extraction and learning features in the chronological data, the MD-GRU is used to simulate the dynamic changes in the historical energy sequence data. The CNN module processes spatiotemporal matrices and plots them into distinct vectors by taking global values rather than max or average. The model predicts the load of power consumption 10 minutes in advance using several statistical techniques concerning predicted and actual values. These predicted values can be used to make wiser judgments and enhance the overall effectiveness of the system. The following are the main contributions of this paper:

    1-A novel multi-directional GRU (MD-GRU) model is presented for the detection of dynamic changes in historical load and energy sequence data

    2-This work integrates MD-GRU and CNN having global average pooling (GAP) is used to estimate the load and energy demand to capture long-term dependencies.

    3-The proposed MD-GRU-CNN model is compared with the baseline models such as LSTM, GRU, Bi-LSTM, Bi-GRU, and CNN models

    4-The proposed model is also assessed using other metrics including mean percentage absolute error (MAPE), root mean square error (RMSE), and conventional assessment measures.

    The rest of the paper is structured as follows: Section 2 describes related work. Sections 3 and 4 contain the deep learning techniques for load and energy forecasting as well as the proposed methodology, while Section 5 contains the experimentation. Section 6 is for discussion, and Section 7 concludes the paper.

    For the load and energy management system, there has been a major growth in the use of modern metering and sensors, which has greatly increased the information. A significant amount of data is produced, serving as a trustworthy source of information for load forecasting algorithms. There are parametric and non-parametric versions of these load forecasting models. While ANNs are among the most widely used non-parametric techniques for creating new energy demand predictions, parametric methodologies are based on analytical models [35,36]. There have been numerous electrical load forecasting models proposed in recent years using the latest deep learning techniques; some of the more popular models that researchers now employ are detailed in [37,38,39].

    To forecast energy consumption over various time horizons, time-series approaches have been frequently used. The most prevalent model is the autoregressive moving average (ARMA) model, which combines the moving average (MA) and autoregressive (AR) models. The AR Model relies on the linear regression of actual values depending on a database, while the MA Model makes utilization of the moving average to compute Regressions employing forecast errors for previous values [40]. Another study uses autoregressive integrated moving average (ARIMA) models to create novel forecasting systems [41]. However, there are numerous intrinsic nonlinear properties in the enormous data, and classic statistical approaches cannot effectively learn this nonlinear data even though these methods may, to some extent, estimate short-term demand. Therefore, it is exceedingly difficult for conventional statistical approaches to forecast load with any degree of accuracy.

    Various machine learning (ML) techniques, such as artificial neural networks fuzzy systems [42] and support vector machines (SVM) [43] have been effectively implemented in predicting and classifying issues [44,45,46]. The development of a system for load and energy forecasting using SVM and ML techniques is effective in addressing the difficulties of nonlinearity [47]. An adaptive neuro-fuzzy inference system (ANFIS), was used in another project to create a short-term load and energy forecasting model [48]. The backpropagation neural network (BPNN) has also been taken into consideration for load forecasting as deep learning has advanced [49]. A hybrid model of BPNN and the multi-label algorithm based on K-nearest neighbor (KNN) and K-means was presented; however, because of the nature of feedforward neural networks, it is unable to effectively learn time-sequence data for the power system [50].

    Recently, computer vision, natural language processing and speech recognition have all benefited greatly from recent advancements in neural networks, especially deep neural networks. Deep neural network applications for short-term load forecasting are currently a novel and widely discussed topic. An analysis of artificial neural networks (ANNs) for predicting short-term load is shown in [51]. Recurrent neural networks (RNN), convolutional neural networks (CNN), and their combination and long short-term memory (LSTM) are just a few of the various building blocks that have made deep neural networks highly adaptable and successful [52].

    A novel approach with LSTM and genetic algorithm (GA) integration was put out in [53], and it produced a mean absolute percentage error with a minimal value. An innovative method for the planning and operation of energy generation called the Bi-GRU (gated recurrent unit) prediction system was created [54]. The bidirectional characteristic of GRU, in contrast to other approaches, is not enough to gather long-term sequential information from all areas of energy and load forecasting. By using positive and negative input, the bidirectional model trains the nodes to operate in forward and reverse modes. The advantage of the bidirectional ability to focus in both directions (forward and backward), as indicated is useful for extracting more information from the long-term information sequence, but it is insufficient for energy and load forecasting [55]. Multiple load and energy forecasters have been produced in which the forecaster's framework must get more intricate to produce forecasting accuracy.

    Sequential information use is carried out via recurrent neural networks. These networks connect words with particular time steps in the sequence. [56,57]. The amount of time steps determines the maximum input sequence length in the final result. Figure 1 depicts the recurrent neural network's architecture.

    Figure 1.  Recurrent neural network architecture.

    The following are the sequential data as feature vectors that are provided to the recurrent neural network (RNN) as inputs after changing feeds:

    hst=f(X.It+Wmhst1+b). (1)
    Ot=softmax(Y.hst). (2)

    The outputs can be observed in the above image by the corresponding time step, however depending on the task, this may not be essential. Our objective for managing the energy and load is to concentrate on the output. A recurrent neural network can accumulate sequential information in the same situation when there is no necessity for input with the appropriate time step and hidden state. Previously, researchers discovered that by introducing LSTM, Bi-LSTM GRU, and Bi-GRU, the recurrent neural network can be improved for learning long term dependencies. However, the bidirectionality was insufficient in the case of load and energy management to achieve the desired outcomes [58].

    Long-term dependency issues are addressed with the LSTM [59]. When a long short-term memory network is trained to use backpropagation with time, it can address the vanishing gradient problem [60]. Although LSTMs are considered to be an adequate methodology, still architecture only considers the past, which can sometimes result in poor performance, particularly in load and energy management where the future data is also crucial. With two hidden states, forward hst and backward hst, bidirectionality in LSTMs can capture both contexts: prior and upcoming.

    LSTM optimization known as the gated recurrent unit (GRU) performs better than conventional LSTMs. When compared to LSTMs, the gated recurrent unit (GRU) has fewer parameters since it consolidates the input gates and forget gates provided in [61] into only two gates. Similar to LSTMs, but faster than Bi-LSTMs, is the bidirectional gated recurrent unit. The Bi-GRU unit can store sequences of information in both directions, such as prior and upcoming information, and have a long short-term memory for preserving long-term dependencies. Many models have been successful in terms of the informational flow relating in to both directions for load and energy management [62]. The output of a bidirectional recurrent module that processes sequences in opposite directions is, for example, the hidden state generated as:

    [hs1hs1,hs2hs2,.....,hsnhsn]. (3)

    Convolutional neural networks were initially for image classification but a short while later, evolved into models that were used for a variety of applications. The sequential data is fed into a convolutional layer, which then generates various filters, learns several features, and applies these features successively to the various input divisions. The output is typically then aggregated to lower dimensions and routed into a linked layer [15]. Multiple features are used in this layer to extract local features. Figure 2 shows the traditional convolutional neural network architecture as it is described in [63].

    Figure 2.  Traditional convolutional neural network architecture.

    To choose a maximal value, the max pooling layer must first discard any other values that are not maximal. A concept of dropout layer normalization is essential to build a dense connection in both the max-pooling and convolutional layers as presented in [64].

    The output of a convolutional layer is processed by a global average pooling layer, as a regularizer, replacing the fully connected layer and dropout. Further, it specifically coordinates between the mapping of features and classifications to examine the computational complexity and determine whether it is less or not. Additionally, no parameter modification is required in the global average pooling [65]. Each feature map's average value produces a unique value, which is then directly entered into the SoftMax algorithm to obtain the likelihood distribution along with confidence values for each class. The following expression illustrates the likelihood of appropriation:

    P(ij|k,θ)=exp(xj(k,θ))1i|X|exp(xi(k,θ)),1j|X|. (4)

    For load and energy forecasting, deep learning techniques, particularly sequential models like RNN and LSTM models, are being used. Feedforward neural networks with internal memory are known as recurrent neural networks. The outcome of the current input depends on the computation of the prior inputs, as opposed to RNNs, which perform the same task for every data input. The output is produced, duplicated, and then distributed throughout the network. However, the issue of gradient vanishing and exploding prevents the typical RNN design from supporting long sequences. A modified version of recurrent neural networks makes it simpler to store past knowledge in memory like long short-term memory (LSTM) and gated recurrent units (GRUs) [66].

    Categorical values require label encoding approaches since deep learning models are unable to understand them. These methods change these category values into numerical ones that are appropriate for deep-learning neural network models. Understanding how the extended input sequence is created, however, is crucial. The time period of the date, such as a certain day, month, etc., is the input data for energy and load management. As a result, it is necessary to transform this kind of data into a sequence that can be used as additional input in the architecture for load and energy forecasting.

    The fundamental principle of MD-GRU is to use as many recurrent connections as there are dimensions in the data to replace the one recurrent connection present in regular RNNs. The hidden layer of the network receives an external input and its activations from one step back along all dimensions at each point in the data sequence during the forward pass. This module has multiple dimensions that can individually scan all four directions in each column and row where hidden layers are calculated in the vertical and horizontal direction at every stage and added up at the end. Two output attribute maps are employed, one for the vertical direction and the other for the horizontal direction. The forward hidden states, backward hidden states, forward update gates, and backward update gates are the four sets of activations that the MD-GRU computes for each time step. The final output of the module is generated by combining these activations, which are calculated using a set of learned weights and biases. The current simulation makes use of the total of these output attribute maps as an alternative to concatenate the output feature maps as shown in Figure 3 below:

    Figure 3.  The architecture of multi-directional gated recurrent unit (MD-GRU).

    A novel hybrid Multi-Directional MD-GRU-CNN is proposed by using the benefits of the combination of MD-GRU for the processing of time sequence data with CNN to handle high dimensionality of data. The spatial and temporal dependencies in the input data are intended to be captured by this hybrid architecture, making it particularly ideal for tasks that call for both short-term and long-term patterns.

    Figure 4 illustrates the proposed hybrid multidimensional MD-GRU-CNN neural network design. With spatiotemporal matrices gathered from the power system as inputs and sequential data based on time as outputs, the proposed structure can forecast the load and energy level in a relatively very short amount of time. The MD-GRU module aims to capture long-term dependency. The MD-GRU module can learn useful information by scanning each column and row independently in different directions in the historical data for a long period of time through the memory cell, and the forget gate will forget the useless information. The inputs to the MD-GRU module are time sequence data. The MD-GRU module has a sizable number of gated recurrent units that are connected CNNs that make use of fully connected layers and global average pooling. While the MD-GRU can identify temporal patterns in the data, such as daily or weekly cycles in energy consumption or load, the CNN can identify spatial aspects of the data, like patterns in the distribution of load or energy consumption throughout an area. Finally, by calculating the mean value of each neuron in the fully connected layers, the load-predicting results may be determined as shown below in the Figure 4 below:

    Figure 4.  Proposed architecture based on hybridization of multi-directional GRU and CNN.

    To evaluate our proposed methodologies, we run the system over the following datasets:

    The UCI machine learning repository is where the first dataset for the experimentation was gathered [67]. The dataset contains 10,000 instances having 14 variables relating to the values of electricity producers, produced, price elasticity factor, the maximum value of equation root and system reliability as described refers to DS-1.

    The second dataset called the Almanac of Minutely Power Dataset, contains different measurements such as (natural gas, electricity, and water) taken by a single household of 19 devices over the duration of an entire year, at a resolution of one (1) minute that is then down-sampled to a resolution of thirty (30) minutes [68] refers to DS-2.

    The third dataset, referred to as the smart grid smart city was started in an initiative by Australian government [69] by collecting smart-meter data of 78,000 clients referred to as DS-3.

    Multiple metrics such as accuracy, precision, recall, and F1 Score are broadly applicable and significant for evaluating the performance of different prediction models. Additionally, the mean absolute percentage error (MAPE) and root mean square error (RMSE) are also added:

    MAPE=1Nni=1|y(i)y(i)y(i)|×100%, (5)
    RMSE=xi=1(ˆγiy)2x (6)

    where n refers to samples training size y(i) actual and y(i) are expected values, respectively.

    The average ratio between the error and actual values at one forecasting point is represented by the MAPE. The sample standard deviation of variations in predicted and actual value as RMSE. The performance of the model's prediction increases with decreasing MAPE and RMSE values.

    This part begins by presenting baseline methods, which consist of LSTM, bidirectional LSTM, GRU, bidirectional GRU, and CNN experiments with various time stamps. Then, by combining Multi-directional GRU with CNN and examining several variations, such as MD-GRU separately and MD-GRU-CNN, as shown in Tables 1 and 2, we demonstrate how we expand these baseline models toward the proposed architecture.

    Table 1.  Baseline models experimented with different timestamps (2, 4, 6).
    Models Narrative
    CNN Baseline Methods with Different Time Stamps
    LSTM
    GRU
    Bi-LSTM
    Bi-GRU

     | Show Table
    DownLoad: CSV
    Table 2.  Proposed models experimented with different timestamps (4, 8, and 12).
    Models Narrative
    MD-GRU Proposed Methods with Different Time Stamps
    MD-GRU-CNN

     | Show Table
    DownLoad: CSV

    The purpose of the testing in this work is to show the effectiveness of the recommended procedure for diversifications in different datasets, in comparison with the tuning of the multiple parameters. During our initial testing, we checked the epoch, batch size, filter windows, MAPE, and learning rate to determine the optimum parameter. However, this work does not go into in-depth detail concerning adjusting hyperparameters.

    With a neural network, we can divide the huge dataset into smaller chunks by training with small batches of several training cases. Accuracy and layers are sometimes directly associated, implying that as the number of layers increases, so does accuracy. Most of the time, this tends to increase precision; but, as the number of layers increases, accuracy can substantially decline due to the "Vanishing Gradient Problem." [38]. Sometimes the computations also rely on how far the single hidden layer is expanded. For instance, if there are not enough hidden layers, the performance will suffer. The dimensions associated with the hidden state are taken into consideration to be 128 and one layer of convolution layers is used in this work along with a field size of (3, 3, and 5).

    After evaluating our model and experimenting with other window filter combinations, we found that seven (7) is the best alternative. Additionally, the batch size is set at 128, the learning rate is set at 0.02 and the epochs maintain within 10 and 40 including all datasets. Dropout, however, was fixed at 0.5 for regularization in the baseline models. Following a series of studies, it is also discovered that the activation function hyperbolic ReLu is perfect in the convolutional layer for the best performance. Lastly, MAPE 0:95 was found appropriate.

    For experimentation, a diversity of evaluation techniques is utilized, including accuracy, precision, recall, and F1-score for both baseline and proposed models to replace the stacking of multiple convolutional layers as well as the problems associated with bidirectional architecture. Average MAPE and RMSE values were also employed to support the findings. Further, several of the most popular neural network models, including CNN, LSTM, Bi-LSTM, GRU and Bi-GRU models were selected as baseline models to assess the effectiveness and robustness of the proposed MD-GRU-CNN model.

    The features of the baseline and (proposed) MD-GRU-CNN models were acquired by learning the training datasets. The model was then validated using test datasets. Tables 36 contain the results considering the above evaluation metrics on multiple datasets referring to DS-1, 2, and DS-3. For baseline models, the most elevated accuracy is 90.99 % achieved Bi-GRU model using DS-3 can be seen in Figure 5. The most appropriate precision and recall also observed on similar models are 91.86% and 93.18% respectively as visualized in Figures 6 and 7. Whereas the F1-Score was 92.97% using DS-1 on the Bi-GRU model as depicted in Figure 8.

    Table 3.  Accuracy achieved on baseline models.
    Accuracy
    Models DS-1 DS-2 DS-3
    CNN 82.4 85.7 84.6
    LSTM 83.8 87.1 87.9
    GRU 86.99 88.12 88.71
    Bi-LSTM 87.8 89.27 90
    Bi-GRU 89.67 92.8 90.99

     | Show Table
    DownLoad: CSV
    Table 4.  Precision on baseline models.
    Models DS-1 DS-2 DS-3
    CNN 83.5 88.7 85.6
    LSTM 85.8 90.21 87.9
    GRU 89.77 90.97 89.45
    Bi-LSTM 90 89.87 90.78
    Bi-GRU 91.27 89.76 91.86

     | Show Table
    DownLoad: CSV
    Table 5.  Recall baseline models.
    Recall
    Models DS-1 DS-2 DS-3
    CNN 85.3 87.6 86.5
    LSTM 88.2 92.1 89.2
    GRU 87.87 91.75 90.25
    Bi-LSTM 92.8 90.71 90.02
    Bi-GRU 93.17 87.06 93.18

     | Show Table
    DownLoad: CSV
    Table 6.  F1-Score on baseline models.
    F1-Score
    Models DS-1 DS-2 DS-3
    CNN 86.2 88.6 83.5
    LSTM 90.2 93.4 88.4
    GRU 89.71 92.15 91.05
    Bi-LSTM 91.99 89.84 91.72
    Bi-GRU 92.97 89.62 92.08

     | Show Table
    DownLoad: CSV
    Figure 5.  Accuracy on baseline models using DS-1, DS-2 and DS-3.
    Figure 6.  Precision on baseline models using DS-1, DS-2 and DS-3.
    Figure 7.  Recall on baseline models using DS-1, DS-2 and DS-3.
    Figure 8.  F1-Score on baseline models using DS-1, DS-2 and DS-3.

    Similarly, the average MAPE attained on the Baseline method is also presented in Tables 79 using all three data sets (DS-1, DS-2, and DS-3) incorporating different time stamps (2, 4, and 6).

    Table 7.  Average MAPE on baseline model using DS-1.
    DS-1 (Average MAPE)
    Models 2-Timestamps 4-Timestamps 6-Timestamps
    CNN 40.5 39.5 41.5
    LSTM 39.8 42.8 40.5
    GRU 38.7 40.71 40.99
    Bi-LSTM 37.8 39.1 38.99
    Bi-GRU 37.9 35.4 35.01

     | Show Table
    DownLoad: CSV
    Table 8.  Average MAPE on baseline model using DS-2.
    DS-2 (Average MAPE)
    Models 2-Timestamps 4-Timestamps 6-Timestamps
    CNN 39.6 41.8 41.53
    LSTM 38.41 40.3 38.55
    GRU 38.34 39.91 37.56
    Bi-LSTM 37.04 38.78 36.92
    Bi-GRU 35.2 36.8 35.79

     | Show Table
    DownLoad: CSV
    Table 9.  Average MAPE on baseline model using DS-3.
    DS-3 (Average MAPE)
    Models 2-Timestamps 4-Timestamps 6-Timestamps
    CNN 38.6 42.4 41.03
    LSTM 36.97 39.47 40.89
    GRU 37.14 39.47 37.58
    Bi-LSTM 37.04 38.14 39.87
    Bi-GRU 36.2 35.49 35.17

     | Show Table
    DownLoad: CSV

    Similarly, in baseline models, the least observed MAPE value is 35.01 on Bi-GRU using DS-1, 35.4 on 4 timestamps using DS-3, and 35.2 on DS-2 using 2 timestamps can be seen in Figures 911. By considering accuracy, precision, as well as recall and F1-score, Tables 1013 demonstrate the effectiveness of the proposed models. The proposed model outperforms by achieving reliable 95.76% accuracy on MD-GRU and 97.85% using DS-3 compared to baseline models depicted in Figure 12.

    Figure 9.  Average MAPE on baseline model using DS-1 with different timestamps.
    Figure 10.  Average MAPE on baseline model using DS-2 with different timestamps.
    Figure 11.  Average MAPE on baseline model using DS-3 with different timestamps.
    Table 10.  Accuracy achieved on proposed models using DS 1, DS-2 and DS-3.
    Accuracy
    Models DS-1 DS-2 DS-3
    MD-GRU 95.21 94.77 95.76
    MD-GRU-CNN 96.17 95.89 97.85

     | Show Table
    DownLoad: CSV
    Table 11.  Precision on proposed models using DS-1, DS-2 and DS-3.
    Precision
    Models DS-1 DS-2 DS-3
    MD-GRU 96.11 95.83 95.61
    MD-GRU-CNN 97.27 96.99 98.55

     | Show Table
    DownLoad: CSV
    Table 12.  Recall on proposed models using DS-1, DS-2 and DS-3.
    Recall
    Models DS-1 DS-2 DS-3
    MD-GRU 98.1 98.73 96.17
    MD-GRU-CNN 96.83 98.99 98.45

     | Show Table
    DownLoad: CSV
    Table 13.  F1-Score on proposed models using DS-1, DS-2 and DS-3.
    F1-Score
    Models DS-1 DS-2 DS-3
    MD-GRU 97.01 97.83 95.84
    MD-GRU-CNN 97.97 97.49 98.85

     | Show Table
    DownLoad: CSV
    Figure 12.  Accuracy was achieved on MD-GRU and MD-GRU-CNN using DS-1, DS-2 and DS-3.

    Figure 12 illustrates how the proposed models, MD-GRU and MD-GRU-CNN, fared better in terms of precision than earlier neural network baseline models employing DS-2 and DS-3, obtaining 95.83% and 98.55% higher values, respectively. On the DS-2, the greatest recall rates were 98.73% and 98.99%, as shown in Figures 13 and 14. Figure 15 indicates how the proposed model surpassed baseline models by gaining an F1-score on DS-3 that was also higher, at 98.85%.

    Figure 13.  Precision on MD-GRU and MD-GRU-CNN using DS-1, DS-2 and DS-3.
    Figure 14.  Recall on MD-GRU and MD-GRU-CNN using DS-1, DS-2 and DS-3.
    Figure 15.  F1-Score on proposed models using different timestamps.

    The proposed models (MD-GRU and MD-GRU-CNN) perform superior to standard architectures of CNN, LSTM, Bi-LSTM, GRU, and Bi-GRU (baseline), as demonstrated in the above tables. This is because the datasets used in this research are spatiotemporal. With the proposed hybridization (fusion of MD-GRU and CNN), there is no possibility of a loss of related information that usually occurs when flattening spatiotemporal data to dimensional and time sequence data. Resulting in the inability of recurrent and its variations models to learn the features effectively.

    Also, the proposed models have the lowest average MAPE values out of all the baseline models shown in Table 14 and Figure 16. From the three datasets, our proposed hybrid model can effectively learn both the time sequence and spatiotemporal data and extract additional features. For a better assessment, the baseline and proposed models' RMSE representations are listed in Table 15 and illustrated in Figure 17. As shown in the mentioned figure, though the RMSE values of the Bi-GRU and Bi-LSTM models are nearly identical, the Bi-GRU model surpasses the GRU and Bi-LSTM models. The MD-GRU model values fluctuate between CNN, GRU, Bi-LSTM and MD-GRU-CNN revealing that the proposed model consistently surpasses the baseline models, while MD-GRU is lower in performance than the MD-GRU-CNN model.

    Table 14.  Average MAPE on proposed methods using DS-1, DS-2 and DS-3.
    DS-1, DS-2, DS-3 (Average MAPE)
    Models 4-Timestamps 8-Timestamps 12-Timestamps
    MD-GRU 26.9 28.93 26.74
    25.2 26.1 25.6
    25.87 26.99 27.12
    MD-GRU-CNN 25.11 25.07 26.21
    26.17 25.18 26
    25.09 25.9 26.14

     | Show Table
    DownLoad: CSV
    Figure 16.  Average MAPE on proposed methods using DS-1, DS-2 and DS-3 with 4, 8 and 12 timestamps.
    Table 15.  RMSE on baseline and proposed models using DS-1, DS-2 and DS-3.
    RMSE
    Models DS-1 DS-2 DS-3
    CNN 2134.45 2013.7 1966.75
    LSTM 1998.5 2010.6 2034.45
    GRU 1962.8 1987.44 2001.69
    Bi-LSTM 1865.7 1871.4 1885.6
    Bi-GRU 1851.4 1849.77 1880.5
    MD-GRU 1598.6 1687.5 1654.36
    MD-GRU-CNN 1418.5 1380.5 1372.96

     | Show Table
    DownLoad: CSV
    Figure 17.  RMSE on baseline and proposed models using DS-1, DS-2 and DS-3 with different timestamps.

    The results presented in Section Ⅳ demonstrate that the proposed methodology consistently outperforms many previous works. Results reveal that the combination of multi-directional gated recurrent unit (MD-GRU) and convolutional neural network enhances the effectiveness of deep learning models. As opposed to this work, which examines a single deep learning architecture across many diverse datasets from various sources referred to as DS-1, DS-2 and DS-3, many earlier efforts typically provide findings on a single dataset. Additionally, it has been established that all sequential models that have bidirectional behavior and multidirectional abilities dominate traditional sequential models including GRU and LSTM. When combining several datasets to forecast load and energy, the integration of MAPE and RMSE is crucial.

    Additionally, it has been found that many machine learning approaches are insufficiently equipped to handle categorical variables. In this scenario, the categorical values should always be transformed into statistical values by using a label-learning algorithm. The initial version of this study presented a multi-directional GRU (MD-GRU) that employs temporal GRUs to autonomously analyze every row and column in different positions. Through each level, hidden layers are figured out and added collectively. In the vertical position, two of the four temporal GRUs are employed. The use of Multidirectional GRUs, as depicted in Figure 2, results from the other Temporal GRUs' use in vertical directions. Because of the computational complexity and parameter optimizations required to learn long-term dependencies, other variants such as LSTM, Bi-LSTM, GRU, and Bi-GRU are insufficient. We conducted multiple experiments to compare the effectiveness of proposed models to baseline models on similar datasets as shown in the above tables using multiple evaluation metrics such as MAPE (Loss Function) and RMSE values.

    Secondly, the integration of two neural networks that utilize the multi-directional GRU-CNN model with global average pooling (GAP) proposed in this research improves the accuracy. The multidirectional GRU (MD-GRU) module is targeted at analyzing the time sequence data to learn long-term dependencies, whereas CNN focuses on processing the spatiotemporal data. The convolution and global average pooling layer are employed effectively in this study to directly extract local features from spatiotemporal data and provide effective representation.

    A global average pooling layer replaces the conventional max-pooling layer that is applied to the output produced by a convolutional layer and also replaces the fully connected layer and dropout since it acts as a regularizer to test whether the computational complexity is reduced or not. Additionally, it precisely correlates between feature mapping and classifications. Rather than including all of the layers of the particular feature map, we emphasized the average value related to each feature map. By separating characteristics from changeable data that affect the power system, the proposed models quickly and correctly forecast energy and load. The MAPE and RMSE values for the MD-GRU-CNN model are the lowest.

    This work has the potential to be expanded in a variety of ways. For each dataset, hyperparameter tuning and optimization can be examined in conjunction with other parameters such as weather conditions, climate changes, and holidays. Simultaneously, to train the deep neural network more effectively as well as to shorten the model's training time new techniques other than MAPE and RSME are required. Finally, data augmentation in the time-series domain and advanced learning techniques can assist to address the issue of data limitation.

    Accurate forecasting of energy and load is essential for improving energy systems and their functionality. These are now producing a large amount of data that is complex, diverse, and of a spatiotemporal nature as a result of digitalization. To learn long-term dependencies linked to sequential data, most classical and deep learning techniques fall short. We introduced two novel deep learning architectures based on consecutive layers that integrate multi-directional (vertical and horizontal) GRU with convolutional neural network employing GAP as hybridization. This fusion uses MD-GRU to extract feature vectorization from time sequence data, whereas CNN extracts high-dimensional data. Multiple datasets were used to train and test the model, and several measures including accuracy, precision, recall and F1-Score, were used to assess its performance. Additionally, the MD-GRU-CNN model produced the lowest MAPE and RMSE values across baseline models for CNN, LSTM, GRU, Bi-LSTM, and Bi-GRU. In the future, similar work can be used for other diverse applications such as image dephasing towards scenically improving the degraded visibility.

    This research was funded by Princess Nourah bint Abdulrahman University and Researchers Supporting Project number (PNURSP2023R346), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

    All authors declare no conflicts of interest in this paper.

    The authors declare they have not used Artificial Intelligence (AI) tools in the creation of this article.



    [1] Trendall AF (1983) The Hamersley Basin, In: Trendall A.F. and Morris R.C. Eds., Iron—formation: facts and problems, Elsevier, 69–129.
    [2] Guo WW (1999) Magnetic petrophysics and density investigations of the Hamersley Province, Western Australia: implications for magnetic and gravity interpretation, Australia, The University of Western Australia.
    [3] Li ZX, Guo W, Powell CMA (2000) Timing and genesis of Hamersley BIF-hosted iron deposits: a new palaeomagnetic interpretation. Perth, Australia, Minerals & Energy Research Institute of Western Australia.
    [4] Bodycoat FM (2010) Stratigraphic and structural setting of iron mineralization at E Deposit (East), Area C, Hamersley Province, Western Australia. Appl Earth Sci 119: 49–55. https://doi.org/10.1179/037174510X12853354810543 doi: 10.1179/037174510X12853354810543
    [5] Geological Survey of Western Australia, Fortescue–Hamersley: Extended abstracts. Perth Australia: Geological Survey of Western Australia, 2022. Available from: https://dmpbookshop.eruditetechnologies.com.au/product/fortescuehamersley-2022-extended-abstracts.do.
    [6] Perring CS, Hronsky JMA, Crowe M (2022) Phanerozoic history of the Pilbara region: implications for iron mineralization. Aust J Earth Sci 69: 757–775. https://doi.org/10.1080/08120099.2022.2048888 doi: 10.1080/08120099.2022.2048888
    [7] Butt AL, Flis MF (1997) The application of geophysics to iron ore mining in the Hamersley Basin, Western Australia. Explor Geophys 28: 195–198. https://doi.org/10.1071/EG997195 doi: 10.1071/EG997195
    [8] Hawke PJ, Flis MF (1997) Application of electrical techniques for iron ore exploration. Exploration Geophysics 28: 242–246. doi: 10.1071/EG997242
    [9] Porath H, Chamalaun FH (1968) Palaeomagnetism of Australian haematite ore bodies, Ⅱ, Western Australia. Geophys J R Astr Soc 15: 253–264. https://doi.org/10.1111/j.1365-246X.1968.tb00184.x doi: 10.1111/j.1365-246X.1968.tb00184.x
    [10] Clark DA, Schmidt PW (1986) Magnetic properties of the banded-iron formations of the Hamersley Group, WA Sydney, Australia, CSIRO Division of Mineral Physics, AMIRA Report 1638.
    [11] Guo WW, Li ZX, Dentith MC (2011) Magnetic petrophysical results from the Hamersley Basin and their implications for interpretation of magnetic surveys. Aust J Earth Sci 58: 317–333. https://doi.org/10.1080/08120099.2011.552984 doi: 10.1080/08120099.2011.552984
    [12] Guo W, Powell CMA, Dentith MC, et al. (1998) Self demagnetisation corrections in magnetic modeling: some examples, Explor Geophys 29: 396–401. https://doi.org/10.1071/EG998396 doi: 10.1071/EG998396
    [13] Guo W, Dentith MC, Bird RT, et al. (2001) Systematic error analysis of demagnetization and implications for magnetic interpretation. Geophysics 66: 562–570. https://doi.org/10.1190/1.1444947 doi: 10.1190/1.1444947
    [14] Trendall AF, Pepper RS (1977) Chemical composition of the Brockman Iron Formation, Western Australia Geological Survey.
    [15] Ewers WE, Morris RC (1980) Chemical and mineralogical data from the uppermost section of the upper BIF member of the Marra Mamba Iron Formation, Commonwealth Scientific and Industrial Research Organization, CSIRO Division of Mineralogy.
    [16] Ewers WE, Morris RC (1981) Studies of the Dales Gorge Member of the Brockman Iron Formation, Western Australia. Econ Geol 76: 1929–1953. https://doi.org/10.2113/gsecongeo.76.7.1929 doi: 10.2113/gsecongeo.76.7.1929
    [17] Gilhome WR (1975) Mount Tom Price iron orebody, Hamersley Iron Province, In: Knight CL, Eds., Economic Geology of Australia and Papua New Guinea, Australas Inst Min Metall Mono 5: 892–898.
    [18] Guo W, Howard D (2000) Interpretation of the crustal structure between the Hamersley and Ashburton Basins from gravity and magnetic data in the Wyloo area, Western Australia. Explor Geophys 31: 33–38. https://doi.org/10.1071/EG00033 doi: 10.1071/EG00033
    [19] Morris RC (1985) Genesis of iron ore in banded iron–formation by supergene and supergene–metamorphic processes—a conceptual model, In: Wolf KH, Eds., Handbook of strata–bound and stratiform ore deposits, Elsevier, 73–235.
    [20] Kneeshaw M (1984) Pilbara iron ore classification—a proposal for a common classification for BIF-derived supergene iron ore. Australas Inst Min Metall 289: 157–162.
    [21] Harmsworth RA, Kneeshaw M, Morris RC, et al. (1990) BIF-derived iron ores of the Hamersley Province. Australas Inst Min Metall Mono 14: 617–642.
    [22] Clout JMF (2006) Iron formation-hosted iron ores in the Hamersley Province of Western Australia. Appl Earth Sci 115: 115–125. https://doi.org/10.1179/174327506X138931 doi: 10.1179/174327506X138931
    [23] White AJR, Legras M, Smith RE, et al. (2014) Deformation-driven, regional-scale metasomatism in the Hamersley Basin, Western Australia. J Metamorphic Geol 32: 417–433. https://doi.org/10.1111/jmg.12078 doi: 10.1111/jmg.12078
    [24] Perring CS (2021) Petrography of martite-goethite ore and implications for ore genesis, South Flank, Hamersley Province, Western Australia. Aust J Earth Sci 68: 782–798. https://doi.org/10.1080/08120099.2021.1863860 doi: 10.1080/08120099.2021.1863860
    [25] Johnson GR, Olhoeft GR (1984) Density of rocks and minerals, In: Carmichael RS, Eds., CRC handbook of physical properties of rocks, Volume Ⅲ, CRC Press Inc.
    [26] Emerson DW (1990) Notes on mass properties of rocks—density, porosity, permeability. Explor Geophy 21: 209–216. https://doi.org/10.1071/EG990209 doi: 10.1071/EG990209
    [27] Price R (1996) Basement structures of the Turner Syncline region, Hamersley Province, Western Australia. The University of Western Australia, Australia.
    [28] Dunlop DJ, Ozdemir O (1997) Rock magnetism, London: Cambridge University Press.
    [29] Aylmer JA, Mathew PJ, Wylie AW (1978) Bulk density of stratified iron ores and its relationship to grade and porosity. Australas Inst Min Metall 265: 9–17.
    [30] Schon JH (1996) Physical properties of rocks: fundamentals and principles of petrophysics, Pergamon.
    [31] Srigutomo W, Heriyanto M, Aufa MH (2019) Gravity inversion of Talwani model using very fast simulated annealing. J Math Fund Sci 51: 177–190. https://doi.org/10.5614/j.math.fund.sci.2019.51.2.7 doi: 10.5614/j.math.fund.sci.2019.51.2.7
    [32] Rao K, Biswas A (2021) Modeling and uncertainty estimation of gravity anomaly over 2D fault using very fast simulated annealing global optimization. Acta Geophys 69: 1735–1751. https://doi.org/10.1007/s11600-021-00649-8 doi: 10.1007/s11600-021-00649-8
    [33] Li MM, Guo W, Verma B, et al. (2009) Intelligent methods for solving inverse problems of backscattering spectra with noise: a comparison between neural networks and simulated annealing. Neural Comput Applic 18: 423–430. https://doi.org/10.1007/s00521-008-0219-x doi: 10.1007/s00521-008-0219-x
    [34] Bailer-Jones DM, Bailer-Jones CAL (2002) Modeling data: Analogies in neural networks, simulated annealing and genetic algorithms. Model-Based Reasoning, Boston, MA: Springer. https://doi.org/10.1007/978-1-4615-0605-8_9
  • This article has been cited by:

    1. Md Babul Islam, Antonio Guerrieri, Raffaele Gravina, Giancarlo Fortino, A Meta-Survey on Intelligent Energy-Efficient Buildings, 2024, 8, 2504-2289, 83, 10.3390/bdcc8080083
    2. Chenmin Ni, Muhammad Fadhil Marsani, Fam Pei Shan, Xiaopeng Zou, Flood prediction with optimized gated recurrent unit-temporal convolutional network and improved KDE error estimation, 2024, 9, 2473-6988, 14681, 10.3934/math.2024714
    3. Francesca Villano, Gerardo Maria Mauro, Alessia Pedace, A Review on Machine/Deep Learning Techniques Applied to Building Energy Simulation, Optimization and Management, 2024, 4, 2673-7264, 100, 10.3390/thermo4010008
    4. Shuguang Li, Kashif Ali, Salem Algarni, Talal Alqahtani, Sohail Ahmad, Fayza Abdel Aziz ElSeabee, Hameed Ullah, Wasim Jamshed, Kashif Irshad, Numerical analysis of thermophoretic particle deposition in a magneto-Marangoni convective dusty tangent hyperbolic nanofluid flow – Thermal and magnetic features, 2024, 13, 2191-9097, 10.1515/ntrev-2023-0190
    5. Salvatore Mancha Gonzales, Hasnain Iftikhar, Javier Linkolk López-Gonzales, Analysis and forecasting of electricity prices using an improved time series ensemble approach: an application to the Peruvian electricity market, 2024, 9, 2473-6988, 21952, 10.3934/math.20241067
    6. Zheng Grace Ma, Nicolai Bo Vanting, Bo Nørregaard Jørgensen, 2024, Residential Heating Demand Prediction Using Time-Series Clustering and Deep Learning, 979-8-3315-1013-8, 1, 10.1109/ICSAI65059.2024.10893794
    7. Xiaojun Wang, Linrong Wang, Nan Liang, Machine learning-driven power demand forecasting models for optimized power management, 2025, 0948-7921, 10.1007/s00202-025-03140-5
    8. Chingfei Luo, Chenzi Liu, Chen Huang, Meilan Qiu, Dewang Li, ARIMA Markov Model and Its Application of China’s Total Energy Consumption, 2025, 18, 1996-1073, 2914, 10.3390/en18112914
    9. Simeng Zhang, Houqing Zhang, Yangcheng Hu, 2025, Statistical Forecasting of Short-term Electricity Load Using a Meteorological Feature-driven Deep Learning Fusion Framework, 979-8-3315-2184-4, 1067, 10.1109/CEEPE64987.2025.11034044
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(4171) PDF downloads(247) Cited by(1)

Figures and Tables

Figures(6)  /  Tables(4)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog