Loading [MathJax]/jax/element/mml/optable/BasicLatin.js
Research article Special Issues

CNN-Trans-SPP: A small Transformer with CNN for stock price prediction

  • Understanding the patterns of financial activities and predicting their evolution and changes has always been a significant challenge in the field of behavioral finance. Stock price prediction is particularly difficult due to the inherent complexity and stochastic nature of the stock market. Deep learning models offer a more robust solution to nonlinear problems compared to traditional algorithms. In this paper, we propose a simple yet effective fusion model that leverages the strengths of both transformers and convolutional neural networks (CNNs). The CNN component is employed to extract local features, while the Transformer component captures temporal dependencies. To validate the effectiveness of the proposed approach, we conducted experiments on four stocks representing different sectors, including finance, technology, industry, and agriculture. We performed both single-step and multi-step predictions. The experimental results demonstrate that our method significantly improves prediction accuracy, reducing error rates by 45%, 32%, and 36.8% compared to long short-term memory(LSTM), attention-based LSTM, and transformer models.

    Citation: Ying Li, Xiangrong Wang, Yanhui Guo. CNN-Trans-SPP: A small Transformer with CNN for stock price prediction[J]. Electronic Research Archive, 2024, 32(12): 6717-6732. doi: 10.3934/era.2024314

    Related Papers:

    [1] Yuanqiang Chen, Jihui Zheng, Jing An . A Legendre spectral method based on a hybrid format and its error estimation for fourth-order eigenvalue problems. AIMS Mathematics, 2024, 9(3): 7570-7588. doi: 10.3934/math.2024367
    [2] Hui-qing Liao, Ying Fu, He-ping Ma . A space-time spectral method for the 1-D Maxwell equation. AIMS Mathematics, 2021, 6(7): 7649-7668. doi: 10.3934/math.2021444
    [3] A.S. Hendy, R.H. De Staelen, A.A. Aldraiweesh, M.A. Zaky . High order approximation scheme for a fractional order coupled system describing the dynamics of rotating two-component Bose-Einstein condensates. AIMS Mathematics, 2023, 8(10): 22766-22788. doi: 10.3934/math.20231160
    [4] Zunyuan Hu, Can Li, Shimin Guo . Fast finite difference/Legendre spectral collocation approximations for a tempered time-fractional diffusion equation. AIMS Mathematics, 2024, 9(12): 34647-34673. doi: 10.3934/math.20241650
    [5] A. K. Omran, V. G. Pimenov . High-order numerical algorithm for fractional-order nonlinear diffusion equations with a time delay effect. AIMS Mathematics, 2023, 8(4): 7672-7694. doi: 10.3934/math.2023385
    [6] Jeong-Kweon Seo, Byeong-Chun Shin . Reduced-order modeling using the frequency-domain method for parabolic partial differential equations. AIMS Mathematics, 2023, 8(7): 15255-15268. doi: 10.3934/math.2023779
    [7] Chuanhua Wu, Ziqiang Wang . The spectral collocation method for solving a fractional integro-differential equation. AIMS Mathematics, 2022, 7(6): 9577-9587. doi: 10.3934/math.2022532
    [8] Mahmoud A. Zaky, Weam G. Alharbi, Marwa M. Alzubaidi, R.T. Matoog . A Legendre tau approach for high-order pantograph Volterra-Fredholm integro-differential equations. AIMS Mathematics, 2025, 10(3): 7067-7085. doi: 10.3934/math.2025322
    [9] Bo Tang, Huasheng Wang . The a posteriori error estimate in fractional differential equations using generalized Jacobi functions. AIMS Mathematics, 2023, 8(12): 29017-29041. doi: 10.3934/math.20231486
    [10] Yingchao Zhang, Yingzhen Lin . An ε-approximation solution of time-fractional diffusion equations based on Legendre polynomials. AIMS Mathematics, 2024, 9(6): 16773-16789. doi: 10.3934/math.2024813
  • Understanding the patterns of financial activities and predicting their evolution and changes has always been a significant challenge in the field of behavioral finance. Stock price prediction is particularly difficult due to the inherent complexity and stochastic nature of the stock market. Deep learning models offer a more robust solution to nonlinear problems compared to traditional algorithms. In this paper, we propose a simple yet effective fusion model that leverages the strengths of both transformers and convolutional neural networks (CNNs). The CNN component is employed to extract local features, while the Transformer component captures temporal dependencies. To validate the effectiveness of the proposed approach, we conducted experiments on four stocks representing different sectors, including finance, technology, industry, and agriculture. We performed both single-step and multi-step predictions. The experimental results demonstrate that our method significantly improves prediction accuracy, reducing error rates by 45%, 32%, and 36.8% compared to long short-term memory(LSTM), attention-based LSTM, and transformer models.



    The Sobolev equation plays an important role in partial differential equations (PDEs) because of its significant physical background, such as consolidation of clay [1], thermodynamics [2] and flow of fluids through fissured rock [3].

    In this article, a Legendre-tau-Galerkin method in time and its multi-interval form will be considered for the following 2D Sobolev equations:

    {tu(x,y,t)εtΔu(x,y,t)μΔu(x,y,t)+γu(x,y,t)=f(x,y,t),(x,y,t)Σ:=Ω×I,u(x,y,1)=u0(x,y),(x,y)ˉΩ,u(x,y,t)=0,(x,y,t)Ω×I, (1.1)

    where coefficients μ, ε, γ are known positive parameters. Though the the time interval is normally (0, T] and T>0, we set the interval as I=(1,1] and the spatial domain is Ω=(1,1)×(1,1). We consider the time interval I=(1,1] just to simplify the presentation of the theoretical analysis and the algorithm implementation process.

    Because of the great difficulty in obtaining analytical solutions of PDEs, various numerical methods [4,5,6,7] have been proposed to approximate the exact solutions. For the Sobolev equations, there have been many studies investigating the numerical solutions. In [8,9,10], some finite volume element methods were presented in space to solve the two-dimensional Sobolev equations combined with the finite difference schemes in time. The continuous interior penalty finite element method, space-time continuous Galerkin method and finite difference streamline diffusion method were applied in [11,12,13] for solving Sobolev equations with convection-dominated term, respectively. In [14], discontinuous Galerkin scheme in space and Crank-Nicolson scheme in time were considered for approaching the exact solutions of generalized Sobolev equations. In [15], Shi and Sun studied an H1-Galerkin mixed finite element method for solving Sobolev equations and presented the existence, uniqueness and superconvergence results of the discrete scheme. In [16,17], a block-centered finite difference scheme and a time discontinuous Galerkin space-time finite element scheme for nonlinear Sobolev equations were established respectively, and stability and global convergence of the schemes were strictly proved. In [18], the Legendre spectral element method in space combined with the Crank-Nicolson finite difference technique in time were considered. In [19], the nonlinear periodic Sobolev equations were investigated by the Fourier spectral method.

    As is well known, the spectral method is distinguished from other numerical methods by its exponential convergence, and when the spectral method is applied to time-dependent partial differential equations in both space and time (namely, space-time spectral method [20,21,22,23,24,25]), the mismatched accuracy caused by the spectral discretization in space and the finite difference method in time can be solved successfully. In [26], we constructed a space-time Legendre spectral scheme for the linear multi-dimensional Sobolev equations for the first time and the exponential convergence was obtained in both space and time. The main purpose of this paper is to study the multi-interval form of the Legendre space-time fully discrete scheme of two-dimensional Sobolev equations by dividing the time interval. It is worth noting that compared with the single interval method, the multi-interval spectral method [27,28,29] can adopt parallel computation, reduce the scale of the problem effectively and improve the flexibility of the algorithm. Considering the asymmetry of the first order differential operator, the fully discrete scheme is constructed by applying a Legendre-tau-Galerkin method in time based on the Legendre Galerkin method in space. In addition, we still apply the Fourier-like basis functions [30] in space to diagonalize the stiff matrix and the mass matrix simultaneously, which greatly saves the computing time and memory.

    The organization of this article is as follows. In Section 2, we first provide some related notations, then establish the single interval Legendre space-time spectral fully discrete scheme of Eq (1.1) and give the stability analysis and L2(Σ)-error estimates. In Section 3, we divide the time interval and develop the multi-interval Legendre space-time spectral fully discrete scheme of the equations, and then strictly prove the L2(Σ)-error estimates. In Section 4, by using Fourier-like basis functions in space and selecting appropriate basis functions in time, we present the implementation of the multi-interval fully discrete scheme. In Section 5, numerical tests are included to access the efficiency and accuracy of the method. Finally, some conclusions are made in Section 6.

    Throughout the paper, the Sobolev spaces in spatial directions are the standard notations used, namely,

    v(x,y)r,Ω=(|ϵ|rDϵv(x,y)2Ω)12,v(x,y)Hr(Ω), (2.1)

    where ϵ=(ϵ1,ϵ2)(ϵi0 are integers and |ϵ|=ϵ1+ϵ2), Dϵv(x,y)=|ϵ|vxϵ1yϵ2 and r,Ω is denoted by Ω when r=0.

    The temporal direction involves a weighted Sobolev space L2ωα,β(I) endowed with the norm and product

    v(t)2I,ωα,β=(v(t),v(t))I,ωα,β=Iv2ωα,βdt,v(t)L2ωα,β(I), (2.2)

    where the weight function is ωα,β(t)=(1t)α(1+t)β. If α=β=0, the norm I and inner product (,)I are denoted in the space L2(I).

    Furthermore, the weighted space-time Sobolev space L2ωα,β(I;Hr(Ω)) is endowed with the norm

    v(x,y,t)L2ωα,β(I;Hr(Ω))=(Iv(x,y,t)2r,Ωωα,βdt)12,v(x,y,t)L2ωα,β(I;Hr(Ω)), (2.3)

    if r=0, the norm L2ωα,β(I;Hr(Ω)) is denoted by L2ωα,β(I;L2(Ω)); if α=β=0, the norm L2ωα,β(I;Hr(Ω)) is denoted by L2(I;Hr(Ω)); if r=0 and α=β=0, the norm L2ωα,β(I;Hr(Ω)) is denoted by L2(I;L2(Ω))=Σ.

    Let Pι be a space of polynomials of degree ι on [1,1] and L=(M,N), where M and N are a pair of given positive integers. In order to develop the single interval Legendre spectral method in time, we define

    V0N={vPN:v(±1)=0},V0N=V0NV0N,VM={vPM:v(1)=0}. (2.4)

    Then applying the Green's formula, we obtain the following single interval Legendre space-time fully discrete scheme of (1.1): Find uLV0NPM(I) satisfying

    {(tuL,v)Σ+ε(tuL,v)Σ+μ(uL,v)Σ+γ(uL,v)Σ=(f,v)Σ,vV0NVM,uL(x,y,1)=P1Nu0(x,y), (2.5)

    where P1N denote the spatial projection operator and its definition will be given below.

    Firstly, we introduce the definition of spatial projection operator and the corresponding lemma. Next, we present the existence, uniqueness and stability conclusion for the solution of (2.5).

    Definition 2.1. [31] Denote H10(Ω)={vH1(Ω):v|Ω=0}, then the orthogonal projection in space P1N:H10(Ω)V0N is given by

    ((P1Nuu),v)Ω=0,vV0N. (2.6)

    Lemma 2.1. [31] If vH10(Ω)Hr(Ω) and r1, we have

    NP1NvvΩ+(P1Nvv)ΩCN1rvr,Ω. (2.7)

    Theorem 2.1. Assume that u0(x,y)H10(Ω)Hr(Ω)(r1) and fL2(Σ), then the scheme (2.5) has a unique solution uL satisfying

    tuLΣ,ω1,0+uLΣ+uLΣC(u01,Ω+fΣ). (2.8)

    Proof. Taking v=(1t)tuL(V0NVM) and using the integration by parts, we can get for the left-hand side of (2.5)

    (tuL,(1t)tuL)Σ=tuL2Σ,ω1,0, (2.9)
    ε(tuL,(1t)tuL)Σ=ε(tuL,(1t)tuL)Σ=εtuL2Σ,ω1,0, (2.10)
    γ(uL,(1t)tuL)Σ=2γuL(1)2Ωγ(uL,(1t)tuL)Σ+γ(uL,uL)Σ, (2.11)

    namely,

    γ(uL,(1t)tuL)Σ=γuL(1)2Ω+γ2uL2Σ, (2.12)

    similarly,

    μ(uL,(1t)tuL)Σ=μuL(1)2Ω+μ2uL2Σ. (2.13)

    Additionally, by the Cauchy-Schwarz inequality and Young's inequality, the right-hand side of (2.5) can be estimated as

    (f,(1t)tuL)ΣfΣ(1t)tuLΣ12f2Σ+tuL2Σ,ω1,0. (2.14)

    Collecting (2.9)–(2.14) leads to

    tuL2Σ,ω1,0+εtuL2Σ,ω1,0+μ2uL2Σ+γ2uL2ΣγuL(1)2Ω+μuL(1)2Ω+12f2Σ+tuL2Σ,ω1,0, (2.15)

    namely,

    tuLΣ,ω1,0+uLΣ+uLΣC(uL(1)Ω+uL(1)Ω+fΣ). (2.16)

    For initial conditions uL(1) and uL(1) in (2.16), according to Lemma 2.1, we can easily get the following estimate:

    uL(1)Ω+uL(1)Ω=P1Nu0Ω+P1Nu0ΩP1Nu0u0Ω+u0Ω+(P1Nu0u0)Ω+u0ΩCNru0r,Ω+u0Ω+CN1ru0r,Ω+u0ΩCu0r,Ω. (2.17)

    Thus, combining estimations (2.16) and (2.17), we immediately attain the stability conclusion.

    Remark 2.1. From stability conclusion, there exists a zero solution if f=0 and u0=0. In other words, we can easily obtain the existence and uniqueness of uL.

    The purpose of this section is to show an L2(Σ)-error estimate of the single interval Legendre space-time spectral method by applying the dual technique. Now, we first introduce following definition and lemma of the time projection operator which will be covered later..

    Definition 2.2. [28] The orthogonal projection in time ΠM:H1(I)PM(I) is given by

    (ΠMuu,v)I=0,vVM, (2.18)

    and ΠMu(1)=u(1).

    Lemma 2.2. [28] (a) If uHσ(I) and σ1, then

    ΠMuuI,ωl,1CM14(1l)σσtuI,ωσ1,σ1,l=0,1. (2.19)

    (b) If uHσ(I) and σ2, then

    ΠMuuI,ω0,1CM18σσtuI,ωσ2,σ2. (2.20)

    Let U=P1NΠMu. Now we decompose the error into: uLu=(uLU)+(Uu) and denote ˜u=uLU. So according to (2.5) we have

    (t˜u,v)Σ+ε(t˜u,v)Σ+μ(˜u,v)Σ+γ(˜u,v)Σ=(t(uU),v)Σ+ε(t(uU),v)Σ+μ((uU),v)Σ+γ(uU,v)Σ,vV0NVM. (2.21)

    According to the Definitions 2.1 and 2.2, for the right-hand side terms of (2.21) we get

    ε(t(uU),v)Σ=ε(t(uΠMu),v)Σ+ε(t(ΠMuP1NΠMu),v)Σ=ε(t(uΠMu),v)Σ, (2.22)
    μ((uU),v)Σ=μ((uP1Nu),v)Σ+μ((P1NuP1NΠMu),v)Σ=0, (2.23)
    γ(uU,v)Σ=γ(uP1Nu,v)Σ+γ(P1NuP1NΠMu,v)Σ=γ(uP1Nu,v)Σ. (2.24)

    Then (2.21) can be simplified as follows:

    (t˜u,v)Σ+ε(t˜u,v)Σ+μ(˜u,v)Σ+γ(˜u,v)Σ=(t(uU),v)Σ+ε(t(IΠM)u,v)Σ+γ((IP1N)u,v)Σ. (2.25)

    According to the definition of ˜u and ΠM, we note

    ˜u(x,y,1)=uL(x,y,1)P1NΠMu(x,y,1)=P1Nu0(x,y)P1Nu(x,y,1)=0. (2.26)

    Now, we give the L2(Σ)-error estimates of the single interval Legendre space-time spectral scheme.

    Theorem 2.2. Suppose uL and u are the solutions of the scheme (2.5) and problem (1.1), respectively. If u0Hr(Ω)H10(Ω) and uHσ(I;Hr(Ω)H10(Ω)) for integers r1, then:

    (a) For σ1,

    uuLΣC{Nr(uL2(I;Hr(Ω))+uu0L2ω0,1(I;Hr(Ω)))+M14σ(σtuL2ωσ1,σ1(I;L2(Ω))+NrσtuL2ωσ1,σ1(I;Hr(Ω))+σtuL2ωσ1,σ1(I;L2(Ω)))}. (2.27)

    (b) For σ2,

    uuLΣC{Nr(uL2(I;Hr(Ω))+uu0L2ω0,1(I;Hr(Ω)))+M18σ(σtuL2ωσ2,σ2(I;L2(Ω))+NrσtuL2ωσ2,σ2(I;Hr(Ω))+σtuL2ωσ2,σ2(I;L2(Ω)))}. (2.28)

    Proof. In order to use the dual technique to attain the L2(Σ)-error estimates, denote H(I)={vH1(I):v(1)=0}. Then for Eq (2.25) we can write its dual equation: For a given gV0NPM(I), obtain ugV0NVM such that

    (te,ug)Σ+ε(te,ug)Σ+μ(e,ug)Σ+γ(e,ug)Σ=(g,e)Σ,eV0N(PM(I)H(I)). (2.29)

    Firstly, we present the existence and uniqueness of ug. Assuming g=0 and taking e=1t[ω1,0ug]t111sugds in (2.29), then we have

    (te,ug)Σ=(t1t[ω1,0ug],ug)Σ=(ω1,0ug,ug)Σ=ug2Σ,ω1,0, (2.30)
    ε(te,ug)Σ=ε(t1t[ω1,0ug],ug)Σ=ε(t1t[ω1,0ug],ug)Σ=ε(ω1,0ug,ug)Σ=εug2Σ,ω1,0, (2.31)
    μ(e,ug)Σ=μ(1t[ω1,0ug],ug)Σ=μ(φ,(1t)φt)Σ=μΩ((1t)φ2|11Iφt((1t)φ)dt)dxdy=μ(φ,t((1t)φ))Σ=μ(φ,(1t)φt))Σ+μ(φ,φ)Σ, (2.32)

    namely,

    μ(e,ug)Σ=μ21t[ω1,0ug]2Σ, (2.33)

    where φ=1t[ω0,1ug] and similarly

    γ(1t[ω1,0ug],ug)Σ=γ21t[ω1,0ug]2Σ. (2.34)

    Collecting (2.30)–(2.34), we have

    ug2Σ,ω1,0+εug2Σ,ω1,0+μ21t[ω1,0ug]2Σ+γ21t[ω1,0ug]2Σ=0. (2.35)

    Then ug=0.

    Now, in order to derive the estimate of ˜u, we first consider the estimates of ug, tug, tug.

    Taking e=(1+t)tug in (2.29), we can get

    (te,ug)Σ=(t((1+t)tug),ug)Σ=Ω((1+t)ugtug)|11dxdy+((1+t)tug,tug)Σ=tug2Σ,ω0,1, (2.36)
    ε(te,ug)Σ=ε(t((1+t)tug),ug)Σ=εtug2Σ,ω0,1, (2.37)
    γ(e,ug)Σ=γ((1+t)tug,ug)Σ=γ(ug,ug)Σ+γ(ug,(1+t)tug)Σ=γ2ug2Σ, (2.38)
    μ(e,ug)Σ=μ((1+t)tug,ug)Σ=μ2ug2Σ. (2.39)

    Collecting (2.36)–(2.39), we can obtain

    tug2Σ,ω0,1+εtug2Σ,ω0,1+γ2ug2Σ+μ2ug2Σ=(g,e)Σ=(g,(1+t)tug)ΣgΣtugΣ,ω0,22gΣtugΣ,ω0,1. (2.40)

    Then we get

    tugΣ,ω0,12gΣ,tugΣ,ω0,12εgΣ,ugΣ2γgΣ. (2.41)

    Next, based on the above estimates we deduce the estimate of ˜u. When g=˜u,e=˜u in (2.29) and utilizing integration by parts, we can see

    ˜u2Σ=(g,˜u)Σ=(t˜u,ug)Σ+ε(t˜u,ug)Σ+μ(˜u,ug)Σ+γ(˜u,ug)Σ=(t(IP1NΠM)u,ug)Σ+ε(t(IΠM)u,ug)Σ+γ((IP1N)u,ug)Σ=((IP1NΠM)(uu0),tug)Σε((IΠM)u,tug)Σ+γ((IP1N)u,ug)Σ(IP1NΠM)(uu0)Σ,ω0,1tugΣ,ω0,1+ε(IΠM)uΣ,ω0,1tugΣ,ω0,1+γ(IP1N)uΣugΣ. (2.42)

    Then we get

    ˜uΣC((IP1NΠM)(uu0)Σ,ω0,1+(IΠM)uΣ,ω0,1+(IP1N)uΣ)C((IP1N)(uu0)Σ,ω0,1+(IΠM)uΣ,ω0,1+(IP1N)(IΠM)uΣ,ω0,1+(IΠM)uΣ,ω0,1+(IP1N)uΣ). (2.43)

    Finally, according to the triangle inequality, we deduce

    uuLΣuUΣ+˜uΣuP1NΠMuΣ+˜uΣC((IP1N)uΣ+(IΠM)uΣ,ω0,1+(IP1N)(IΠM)uΣ,ω0,1+(IP1N)(uu0)Σ,ω0,1+(IΠM)uΣ,ω0,1). (2.44)

    Then, by Lemmas 2.1 and 2.2, we directly derive the final L2(Σ)-error estimates.

    In order to construct the multi-interval form of the Legendre space-time spectral fully discrete scheme, we take a1=1, aK+1=1, ak<ak+1 and denote Ik=(ak,ak+1], ck=ak+1ak, dk=ck/2, namely, I=Kk=1Ik, where K is a known positive integer.

    Denote

    v(t)2I,ωα,β=Ikvk(t)2Ik,˜ωα,β,v(t)L2ωα,β(I), (3.1)

    where vk(t)=v(t)|Ik,˜ωα,β=(1t)αdαk(21tdk)β(tIk) and the definition of vk(t)2Ik,˜ωα,β is presented in Section 2.

    Moreover, let Σk=Ω×Ik and denote

    v(x,y,t)2L2ωα,β(I;Hr(Ω))=Ikvk(x,y,t)2L2˜ωα,β(Ik;Hr(Ω)),v(x,y,t)L2ωα,β(I;Hr(Ω)), (3.2)

    where vk(t)=v(t)|Σk and the definition of vk(x,y,t)L2˜ωα,β(Ik;Hr(Ω)) is presented in Section 2.

    Let M=(M1,,MK) and L=(N,M). Define the space of trial and test functions in time

    XMK=YMKH1(I),YMK={ν:ν|IkPMk(Ik),1kK},˜XMK={ν:ν=(1t)q(t),q(t)YM1K}, (3.3)

    where PMk(Ik) is a space of polynomials of degree Mk on time span Ik and M1=(M11,,MK1).

    Then applying the Green's formula, we can write the multi-interval Legendre space-time spectral fully discrete scheme of (1.1) as: Find uKLV0NXMK satisfying

    {(tuKL,v)Σ+ε(tuKL,v)Σ+μ(uKL,v)Σ+γ(uKL,v)Σ=(f,v)Σ,vV0N˜XMK,uKL(x,y,1)=P1Nu0(x,y). (3.4)

    The stability analysis of (3.4) is similar to the single interval method, so we only provide the error analysis of the scheme. Firstly, we introduce the following definition and lemma of the multi-interval projection operator in time.

    Definition 3.1. [28] The orthogonal projection in time ΠM:H1(I)XMK is given by

    (t(ΠMuu),v)I=0,v˜XMK, (3.5)

    with ΠMu(1)=u(1).

    Lemma 3.1. [28] If uHσ(I) and σ1,ˉM=min1kKMk, we have

    lt(ΠMuu)I,ωl,l1CˉMl1Kk=1(d1kMk)1σσtukIk,l=0,1, (3.6)

    where uk=u|Ik.

    Denote UK=P1NΠMu. Now, we decompose the error into: uuKL=(uUK)+(UKuKL) and denote ˜uK=UKuKL. We note ˜uK(x,y,1)=0. Then according to the scheme (3.4), vV0N˜XMK, we have

    (t˜uK,v)Σ+ε(t˜uK,v)Σ+μ(˜uK,v)Σ+γ(˜uK,v)Σ=(t(UKu),v)Σ+ε(t(UKu),v)Σ+μ((UKu),v)Σ+γ(UKu,v)Σ. (3.7)

    By using the Definitions 2.1 and 3.1, we can see for some terms of the formula (3.7)

    (t(P1NΠMuu),v)Σ=(t(P1NΠMuP1Nu),v)Σ+(t(P1NI)u,v)Σ=(t(P1NI)u,v)Σ, (3.8)
    ε(t(P1NΠMuu),v)Σ=ε(t(ΠMP1NuP1Nu),v)Σ+ε((P1Ntutu,v)Σ=0, (3.9)
    μ((P1NΠMuu),v)Σ=μ((ΠMP1NuΠMu),v)Σ+μ((ΠMI)u,v)Σ=μ((ΠMI)u,v)Σ. (3.10)

    Then (3.7) can be simplified as follows:

    (t˜uK,v)Σ+ε(t˜uK,v)Σ+μ(˜uK,v)Σ+γ(˜uK,v)Σ=(t(P1NI)u,v)Σ+μ((ΠMI)u,v)Σ+γ((P1NΠMI)u,v)Σ,vV0N˜XMK. (3.11)

    Now, we give the L2(Σ)-error estimate of the multi-interval Legendre space-time spectral scheme.

    Theorem 3.1. Suppose u and uKL are the solutions of the problem (1.1) and scheme (3.4), respectively. If uHσ(I;Hr(Ω)H10(Ω)) for integers r,σ1, then

    uuKLΣC{Nr(utL2(I;Hr(Ω))+uL2(I;Hr(Ω)))+ˉM1Kk=1(d1kMk)1σ(σtukL2(Ik;L2(Ω))+σtukL2(Ik;L2(Ω))+NrσtukL2(Ik;Hr(Ω)))}, (3.12)

    where uk=u|Σk.

    If dk=d,Mk=M, we have

    uuKLΣC{Nr(utL2(I;Hr(Ω))+uL2(I;Hr(Ω)))+dMσ(σtuL2(I;L2(Ω))+σtuL2(I;L2(Ω))+NrσtuL2(I;Hr(Ω)))}. (3.13)

    Proof. Similar to the analysis of the L2(Σ)-error estimate of the single interval scheme, we also take the dual technique to give the proof. Considering the dual equation of (3.11), for a given gV0NXMK, we obtain ugV0N˜XMK such that

    (tw,ug)Σ+ε(tw,ug)Σ+μ(w,ug)Σ+γ(w,ug)Σ=(g,w)Σ,wV0N(YMKH(I)). (3.14)

    The existence and uniqueness of ug can be easily attained, so we focus on using the dual equation (3.14) to present the L2(Σ)-error estimate.

    In order to derive the estimate of ˜u, we take g=˜uK and w=˜uK in Eq (3.14),

    ˜uK2Σ=(g,˜uK)Σ=(t˜uK,ug)Σ+ε(t˜uK,ug)Σ+μ(˜uK,ug)Σ+γ(˜uK,ug)Σ=(t(P1NI)u,ug)Σ+μ((ΠMI)u,ug)Σ+γ((P1NΠMI)u,ug)Σ(P1NI)utΣ,ω1,0ugΣ,ω1,0+μ(ΠMI)uΣ,ω1,0ugΣ,ω1,0+γ(P1NΠMI)uΣ,ω1,0ugΣ,ω1,0. (3.15)

    Now, to deduce the estimates of of ug and ug, taking w=t111sugds1t[ω1,0ug] in Eq (3.14),

    ug2Σ,ω1,0+εug2Σ,ω1,0+μ21t[ω1,0ug]2Σ+γ21t[ω1,0ug]2Σ=(g,1t[ω1,0ug])ΣgΣ1t[ω1,0ug]Σ, (3.16)

    then we get

    1t[ω1,0ug]Σ2γgΣ,ugΣ,ω1,02γgΣ,ugΣ,ω1,02γεgΣ. (3.17)

    Taking estimates (3.17) into inequality (3.15) have

    ˜uK2ΣC((IP1N)utΣ,ω1,0gΣ+(ΠMI)uΣ,ω1,0gΣ+(ΠMP1NI)uΣ,ω1,0gΣ), (3.18)

    namely, error estimate of ˜uK is that

    ˜uKΣC((IP1N)utΣ,ω1,0+(ΠMI)uΣ,ω1,0+(ΠMP1NI)uΣ,ω1,0)C((IP1N)utΣ+(ΠMI)uΣ+(ΠMP1NI)uΣ). (3.19)

    Finally, according to the triangle inequality, we deduce

    uuKLΣuUKΣ+˜uKΣC((P1NI)utΣ+(ΠMI)uΣ,ω0,1+(ΠMP1NI)uΣ)C((P1NI)utΣ+(ΠMI)uΣ,ω0,1+(P1NI)uΣ+(ΠMI)uΣ,ω0,1+(P1NI)(ΠMI)uΣ,ω0,1). (3.20)

    In this section, we present the detailed implementation of the multi-interval case by taking Fourier-like basis functions in space and the appropriate basis functions in time. Regarding the single interval case, please see [26] for more information.

    Let

    XMkk={ν:ν=(1t)q(t),q(t)PMk1(Ik),tIk}u(k)(x,y,t):=uKL(x,y,t)|Σk,f(k)(x,y,t):=f(x,y,t)|Σk,1kK. (4.1)

    Then, according to the scheme (3.4), we can see for each k(1kK), we find u(k)V0NPMk(Ik) such that

    {(tu(k),v(k))Σk+ε(tu(k),v(k))Σk+μ(u(k),v(k))Σk+γ(u(k),v(k))Σk=(f(k),v(k))Σk,u(k)(x,y,ak)=u(k1)(x,y,ak),v(k)V0NXMkk, (4.2)

    where u(0)(x,y,a1)=P1Nu0(x,y).

    Furthermore, let

    u(k)(x,y,t)=w(k)(x,y,t)+u(k1)(x,y,ak),

    then the scheme (4.2) can be converted into: Find w(k)V0NPMk(Ik) such that

    {(tw(k),v(k))Σk+ε(tw(k),v(k))Σk+μ(w(k),v(k))Σk+γ(w(k),v(k))Σk=(f(k),v(k))Σkμ(u(k1)(ak),v(k))Σkγ(u(k1)(ak),v(k))Σk,v(k)V0NXMkk,w(k)(x,y,ak)=0. (4.3)

    In order to the operability of the scheme (4.3), we try to separate it into two parts: Find w(k)qV0NPMk(Ik)(q=1,2) such that

    {(tw(k)1,v(k))Σk+ε(tw(k)1,v(k))Σk+μ(w(k)1,v(k))Σk+γ(w(k)1,v(k))Σk=(f(k),v(k))Σkμ(P1Nu0,v(k))Σkγ(P1Nu0,v(k))Σk,v(k)V0NXMkk,w(k)1(x,y,ak)=0, (4.4)
    {(tw(k)2,v(k))Σk+ε(tw(k)2,v(k))Σk+μ(w(k)2,v(k))Σk+γ(w(k)2,v(k))Σk=μ(u(k1)(ak),v(k))Σk+γ(u(k1)(ak),v(k))Σkμ(P1Nu0,v(k))Σkγ(P1Nu0,v(k))Σk,w(k)2(x,y,ak)=0,v(k)V0NXMkk, (4.5)

    The solution of the scheme (4.3) is obtained by

    w(k)(x,y,t)=w(k)1(x,y,t)w(k)2(x,y,t). (4.6)

    The choice of basis functions ϕn(x) and ϕs(y)(0n,sN2) can refer [26]. Regarding the temporal local trail functions ψkm(t) and test functions ˜ψkm(t)(0mMk1), please see [28]. Let

    w(k)(x,y,t)=Mk1m=0N2n,s=0w(k)m,n,sϕn(x)ϕs(y)ψkm(t),W(k)=(w(k)m,n,s)Mk×(N1)2,w(k)q(x,y,t)=Mk1m=0N2n,s=0w(k)q,m,n,sϕn(x)ϕs(y)ψkm(t),W(k)=(w(k)q,m,n,s)Mk×(N1)2,q=1,2, (4.7)

    and v(k)(x,y,t)=ϕn(x)ϕs(y)˜ψ(k)m(t), where n,s=0,1,,N2 and m=0,1,,Mk1, we can see that W(k)=W(k)1W(k)2.

    Denote the sets of LGL points and weights in spatial directions by {x˜n,ϖ˜n}N+1˜n=0 and {y˜s,ϖ˜s}N+1˜s=0, and denote the set of LGL points in time span Ik by {t˜m,˙ϖ˜m}Mk+1˜m=0. Define

    ˜Ψ(k)=(˜ψ(k)m(t˜m))Mk×(Mk+2),F(k)=(f˜m,˜n,˜s)(Mk+2)×(N+2)2,f˜m,˜n,˜s=bk2(f(k)(x˜n,y˜s,t˜m)+μΔP1Nu0(x˜n,y˜s)γP1Nu0(x˜n,y˜s)). (4.8)

    Then, we can get the matrix form of the scheme (4.4) as follows:

    (C(k)+γD(k))W(k)1[Diag(λn)Diag(λs)]+(εC(k)+μD(k))W(k)1[IN1Diag(λn)+Diag(λs)IN1]=˜Ψ(k)Diag(˙ϖ˜m)F(k)[(ΦxDiag(ϖ˜n))T(ΦyDiag(ϖ˜s))T], (4.9)

    where matrices Diag(λn), Diag(λs), IN1, Φx and Φy in spatial directions are given in [26], and matrices C(k) and D(k)(1kK) in temporal direction are given in [28]. Then, by the properties of matrix multiplication in [32], Eq (4.9) can be formulated as

    A(k)vec(W(k)1)=vec(˜Ψ(k)Diag(˙ϖ˜m)F(k)[(ΦxDiag(ϖ˜n))T(ΦyDiag(ϖ˜s))T]), (4.10)

    where

    A(k)=Diag(λn)Diag(λs)(C(k)+γD(k))+[IN1Diag(λs)+Diag(λn)IN1](εC(k)+μD(k)). (4.11)

    Now, we try to get the matrix form of scheme (4.5). According to the processing means in [28], similarly, we have

    (4.12)

    Then

    (4.13)

    where are corresponding eigenvalues of mass matrices and can refer to (4.33) in [28]. Denote , then according to the values of we can get the matrix form of scheme (4.5) as follows:

    (4.14)

    where

    (4.15)

    The Eq (4.14) above can also be converted to a form similar to Eq (4.10).

    In summary, the algorithm can be implemented as follows:

    (1) For each compute by (4.9).

    (2) Obviously, For each assume that have been obtained, then is obtained by (4.14) easily.

    (3) for each

    We mainly devote this section to demonstrate the accuracy and efficiency of the multi-interval Legendre space-time spectral method by utilizing numerical examples for the 2D Sobolev equations. Regarding the numerical results of the single interval Legendre space-time spectral methods for the multi-dimensional Sobolev equations, one can refer [26].

    Example 5.1. We consider the 2D Sobolev equations (1.1) on the time interval with the following exact solution:

    (5.1)

    In this example, the two-interval Legendre space-time spectral method is applied to attain the numerical solution . We divide the time interval into and , namely, and . Under the premise of setting constants , we compare the images of numerical solutions and exact solutions at different times in Figures 1 and 2. From these figures we can deduce that the image of the numerical solutions very well simulate the image of the exact solutions.

    Figure 1.  (Left) The exact solution at time . (Right) The numerical solution by two-interval Legendre spectral method at time for and . Take .
    Figure 2.  (Left) The exact solution at time . (Right) The numerical solution by two-interval Legendre spectral method at time for and . Take .

    To show the spectral accuracy of the proposed method, we plot the maximum point-wise errors and -errors using semi-log coordinates in Figure 3. The numerical results indicate that the proposed method obtained the exponential convergence in both time and space.

    Figure 3.  Spectral accuracy. (Left) Temporal errors by taking . (Right) Spatial errors by taking . Take .

    In Tables 1 and 2, we show -errors in space and time, respectively, mainly to compare the numerical effects of the proposed method applying the Fourier-like basis functions and the traditional basis functions. We can then observe that multi-interval method taking Fourier-like basis functions attain better efficiency.

    Table 1.  Spatial errors. Take and .
    Fourier-like basis Standard basis Order
    CPU(s) CPU(s)
    12 1.4243E-05 0.3309 1.4243E-05 9.7683
    14 1.7056E-07 0.3463 1.7056E-07 26.4171
    16 1.5700E-09 0.3552 1.5700E-09 55.0611
    18 1.1462E-11 0.3734 1.1462E-11 122.2270
    20 3.0801E-13 0.3892 3.0801E-13 237.8722

     | Show Table
    DownLoad: CSV
    Table 2.  Temporal errors. Take and .
    Fourier-like basis Standard basis Order
    CPU(s) CPU(s)
    (10, 10) 8.3000E-03 0.3431 8.3000E-03 4.6588
    (15, 15) 6.2018E-05 0.3495 6.2018E-05 13.9264
    (20, 20) 7.0981E-07 0.3579 7.0981E-07 30.4591
    (25, 25) 8.9560E-09 0.3621 8.9560E-09 61.2795
    (30, 30) 1.0813E-10 0.3785 1.0813E-10 94.7840
    (35, 35) 1.2687E-12 0.3862 1.2687E-12 150.8296

     | Show Table
    DownLoad: CSV

    In Table 3, we compare the temporal -errors obtained by using the single-interval Legendre spectral method and the two-interval Legendre spectral method for the same and . One can find that the two-interval method show a great improvement in accuracy compared with the single-interval method.

    Table 3.  Temporal errors. Take .
    Single interval method Two-interval method
    Order Order
    20 20 2.3710E-01 (10, 10) 8.3000E-03
    20 30 3.4000E-03 (15, 15) 6.2018E-05
    20 40 3.9093E-05 (20, 20) 7.0981E-07
    20 50 3.8171E-07 (25, 25) 8.9560E-09
    20 60 3.2869E-09 (30, 30) 1.0813E-10
    20 70 2.5627E-11 (35, 35) 1.2687E-12

     | Show Table
    DownLoad: CSV

    Example 5.2. In this example we consider the 2D Sobolev equations (1.1) on time interval with the unknown exact solution. The source term is and the initial condition is taken as

    (5.2)

    We also consider the two-interval Legendre space-time spectral method in time. We divide the time interval into and , namely, and . In Figure 4, we depict the numerical solutions at and with and .

    Figure 4.  The numerical solutions by two-interval Legendre space-time spectral method at (Left) and (Right) for and . Take .

    For the unknown of the exact solution, there is no uniform standard to compare the efficiency of the single interval Legendre space-time spectral method with the two-interval method. Thus in Figure 5 by taking the numerical solutions obtained under and as a reference, we plot the temporal and spatial errors of the proposed method with . One can clearly observe that two-interval Legendre space-time spectral method possess the spectral accuracy both in time and space.

    Figure 5.  Spectral accuracy. (Left) Temporal errors by taking . (Right) Spatial errors by taking . Take .

    Example 5.3. This example is devoted to exploring the 2D Sobolev equations (1.1) on time interval with the exact solution, which is not regular enough and unknown in advance. The source term is and the initial condition is taken as, see Figure 6,

    (5.3)
    Figure 6.  Initial function .

    We divide the time interval into and to consider the two-interval Legendre space-time spectral method. In Figure 7, we depict the numerical solutions derived by the proposed method at and with and .

    Figure 7.  The numerical solutions by two-interval Legendre space-time spectral method at (Left) and (Right) for and . Take .

    In Figure 8, by taking the numerical solutions obtained under and as a reference, we present the temporal and spatial errors with . One can clearly observe that our method possess spectral accuracy both in time and space.

    Figure 8.  Spectral accuracy. (Left) Temporal errors by taking . (Right) Spatial errors by taking . Take .

    As previously seen, the spectral method is commonly used to formulate the numerical scheme in space combined with the finite difference method in time, but in most cases, the infinite accuracy in space and finite accuracy in time leads to a unbalanced scheme. In order to avoid this problem, in this paper we use the Legendre-Galerkin method in space and the Legendre-tau-Galerkin method in time to study two-dimensional Sobolev equations. We have succeeded in obtaining spectral convergence in both time and space. In particular, the multi-interval form of the proposed method is also considered. In the theoretical analysis, we not only prove the stability of the single and multi-interval numerical scheme, but also give strict proof of the -error estimates by using the dual technique, where a better error estimate is obtained for the single interval form and the optimal error estimate is obtained for multi-interval form. Compared with the single interval method, the multi-interval spectral method succeeds in reducing the scale of problems, adopting the parallel computers and improving the flexibility of algorithm. Another highlight of this paper is that the Fourier-like basis functions, different from the traditional basis functions, are adopted for the Legendre spectral method in space. Since the mass matrix obtained by Fourier-like basis functions is a diagonal matrix, the computing time and memory can be effectively reduced in the implementation of the algorithm. Numerical experiments show that our method can attain the spectral accuracy both in time and space, and the multi-interval method in time is more efficient than the single one.

    In a future study, we will extend our method to the numerical solutions of the nonlinear Sobolev equation by using appropriate technique to deal with the nonlinear terms effectively.

    This work was supported by the National Natural Science Foundation of China (12161063), Natural Science Foundation of Inner Mongolia Autonomous Regions (2021MS01018), Program for Innovative Research Team in Universities of Inner Mongolia Autonomous Region (NMGIRT2207).

    We declare no conflicts of interest in this paper.



    [1] L. Zhang, F. Wang, B. Xu, W. Chi, Q. Wang, T. Sun, Prediction of stock prices based on LM-BP neural network and the estimation of overfitting point by RDCI, Neural Comput. Appl., 30 (2018), 1425–1444. https://doi.org/10.1007/s00521-017-3296-x doi: 10.1007/s00521-017-3296-x
    [2] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature, 521 (2015), 436–444. https://doi.org/10.1038/nature14539 doi: 10.1038/nature14539
    [3] R. F. Engle, Autoregressive conditional heteroscedasticity with estimates of the variance of United Kingdom inflation, Econometrica, 50 (1982), 987–1007. https://doi.org/10.2307/1912773 doi: 10.2307/1912773
    [4] T. Bollerslev, Generalized autoregressive conditional heteroskedasticity, J. Econom., 31 (1986), 307–327. https://doi.org/10.1016/0304-4076(86)90063-1 doi: 10.1016/0304-4076(86)90063-1
    [5] A. K. Jain, J. Mao, K. M. Mohiuddin, Artificial neural networks: A tutorial, IEEE Comput., 29 (1996), 31–44. https://doi.org/10.1109/2.485891 doi: 10.1109/2.485891
    [6] J. A. K. Suykens, J. Vandewalle, Least squares support vector machine classifiers, Neural Process. Lett., 9 (1999), 293–300. https://doi.org/10.1023/A:1018628609742 doi: 10.1023/A:1018628609742
    [7] F. E. H. Tay, L. Cao, Application of support vector machines in financial time series forecasting, IEEE Comput., 29 (2001), 309–317. https://doi.org/10.1016/S0305-0483(01)00026-3 doi: 10.1016/S0305-0483(01)00026-3
    [8] B. Egeli, M. Ozturan, B. Badur, Stock market prediction using artificial neural networks, Decis. Support Syst., 22 (2003), 171–185.
    [9] Y. Kara, M. A. Boyacioglu, Ö. K. Baykan, Predicting direction of stock price index movement using artificial neural networks and support vector machines: The sample of the Istanbul Stock Exchange, Expert Syst. Appl., 38 (2011), 5311–5319. https://doi.org/10.1016/j.eswa.2010.10.027 doi: 10.1016/j.eswa.2010.10.027
    [10] G. Armano, M. Marchesi, A. Murru, A hybrid genetic-neural architecture for stock indexes forecasting, Inf. Sci., 170 (2005), 3–33. https://doi.org/10.1016/j.ins.2003.03.023 doi: 10.1016/j.ins.2003.03.023
    [11] J. Fu, K. S. Lum, M. N. Nguyen, J. Shi, Stock prediction using fcmac-byy, in Advances in Neural Networks – ISNN 2007, Springer Berlin Heidelberg, 4492 (2007), 346–351. https://doi.org/10.1007/978-3-540-72393-6_42
    [12] R. Choudhry, K. Garg, A hybrid machine learning system for stock market forecasting, Int. J. Comput. Inf. Eng., 2 (2008), 689–692.
    [13] M. Vijh, D. Chandola, V. A. Tikkiwal, A. Kumar, Stock closing price prediction using machine learning techniques, Procedia Comput. Sci., 167 (2020), 599–606. https://doi.org/10.1016/j.procs.2020.03.326 doi: 10.1016/j.procs.2020.03.326
    [14] K. S. Chandar, H. Punjabi, Cat swarm optimization algorithm tuned multilayer perceptron for stock price prediction, Int. J. Web-Based Learn. Teach. Technol., 17 (2022), 1–15. https://doi.org/10.4018/IJWLTT.303113 doi: 10.4018/IJWLTT.303113
    [15] Y. Guo, S. Han, C. Shen, Y. Li, X. Yin, Y. Bai, An adaptive SVR for high-frequency stock price forecasting, IEEE Access, 6 (2018), 11397–11404. https://doi.org/10.1109/ACCESS.2018.2806180 doi: 10.1109/ACCESS.2018.2806180
    [16] B. W. Wanjawa, L. Muchemi, ANN model to predict stock prices at stock exchange markets, preprint, arXiv: 1502.06434.
    [17] A. Tsantekidis, N. Passalis, A. Tefas, J. Kanniainen, M. Gabbouj, A. Iosifidis, Forecasting stock prices from the limit order book using convolutional neural networks, in 2017 IEEE 19th Conference on Business Informatics (CBI), IEEE, (2017), 7–12. https://doi.org/10.1109/CBI.2017.23
    [18] M. U. Gudelek, S. A. Boluk, A. M. Ozbayoglu, A deep learning based stock trading model with 2-D CNN trend detection, in 2017 IEEE Symposium Series on Computational Intelligence (SSCI), IEEE, (2017), 1–8. https://doi.org/10.1109/SSCI.2017.8285188
    [19] A. J. P. Samarawickrama, T. G. I. Fernando, A recurrent neural network approach in predicting daily stock prices an application to the Sri Lankan stock market, in the 2017 IEEE International Conference on Industrial and Information Systems (ICIIS), IEEE, (2017), 1–6. https://doi.org/10.1109/ICIINFS.2017.8300345
    [20] M. Roondiwala, H. Patel, S. Varma, Predicting stock prices using LSTM, Int. J. Sci. Res., 6 (2017), 1754–1756. https://doi.org/10.21275/ART20172755 doi: 10.21275/ART20172755
    [21] S. Selvin, R. Vinayakumar, E. A. Gopalakrishnan, V. K. Menon, K. P. Soman, Stock price prediction using LSTM, RNN and CNN-sliding window model, in 2017 International Conference on Advances in Computing, Communications and Informatics (ICACCI), IEEE, (2017), 1643–1647. https://doi.org/10.1109/ICACCI.2017.8126078
    [22] W. Lu, J. Li, Y. Li, A. Sun, J. Wang, A CNN-LSTM-based model to forecast stock prices, Complexity, 1 (2020), 1–10. https://doi.org/10.1155/2020/6622927 doi: 10.1155/2020/6622927
    [23] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, et al., Attention is all you need, in Advances in Neural Information Processing Systems, 30 (2017).
    [24] Y. Wang, R. Huang, S. Song, Z. Huang, G. Huang, Not all images are worth 16 16 words: dynamic transformers for efficient image recognition, in Advances in Neural Information Processing Systems, 34 (2021), 11960–11973.
    [25] P. Xu, X. Zhu, D. A. Clifton, Multimodal learning with transformers: A survey, IEEE Trans. Pattern Anal. Mach. Intell., 45 (2023), 12113–12132. https://doi.org/10.1109/TPAMI.2023.3275156 doi: 10.1109/TPAMI.2023.3275156
    [26] S. Lai, Mi. Wang, S. Zhao, G. R. Arce, Predicting high-frequency stock movement with differential Transformer neural network, Electronics, 12 (2023), 2943. https://doi.org/10.3390/electronics12132943 doi: 10.3390/electronics12132943
    [27] Z. Tao, W. Wu, J. Wang, Series decomposition Transformer with period-correlation for stock market index prediction, Expert Syst. Appl., 237 (2024), 121424. https://doi.org/10.1016/j.eswa.2023.121424 doi: 10.1016/j.eswa.2023.121424
    [28] A. K. Mishra, J. Renganathan, A. Gupta, Volatility forecasting and assessing risk of financial markets using multi-transformer neural network based architecture, Eng. Appl. Artif. Intell., 133 (2024), 108223. https://doi.org/10.1016/j.engappai.2024.108223 doi: 10.1016/j.engappai.2024.108223
    [29] Z. Shi, MambaStock: Selective state space model for stock prediction, preprint, arXiv: 2402.18959.
    [30] X. Wen, W. Li, Time series prediction based on LSTM-attention-LSTM model, IEEE Access, 11 (2023), 48322–48331. https://doi.org/10.1109/ACCESS.2023.3276628 doi: 10.1109/ACCESS.2023.3276628
    [31] D. O. Oyewola, S. A. Akinwunmi, T. O. Omotehinwa, Deep LSTM and LSTM-Attention Q-learning based reinforcement learning in oil and gas sector prediction, Knowledge-Based Syst., 284 (2024), 111290. https://doi.org/10.1016/j.knosys.2023.111290 doi: 10.1016/j.knosys.2023.111290
  • Reader Comments
  • © 2024 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(1856) PDF downloads(119) Cited by(2)

Figures and Tables

Figures(8)  /  Tables(3)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog