A Multivariate Time Series Prediction Method Based on a Two-Branch Liquid State-Space Model

By using explicit variable decoupling and a liquid core module in a two-branch liquid state-space model, the problems of microscopic rigidity and macroscopic entanglement in multivariate time series prediction models are solved, enabling real-time capture and accurate prediction of transient dynamic characteristics.

CN123070183APending Publication Date: 2026-11-10BEIJING UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610965924.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-01
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing multivariate time series forecasting models suffer from rigid micro-level representations and entangled macro-level features when dealing with highly chaotic, highly volatile environments rich in local mutations, resulting in an inability to capture short-term mutations and severe lag in prediction.

Method used

A method based on a dual-branch liquid state-space model is adopted. Through explicit variable decoupling and a liquid core module, dynamic time-varying features are extracted from the autoregressive branch and the multivariate residual branch, respectively. The Poly-B liquid permeation valve is used for dynamic modulation to generate a feature sequence that dynamically deforms in real time with the input intensity. The predicted values ​​are dynamically fused through a neural gating mechanism, breaking the traditional linear time-invariant constraint.

Benefits of technology

It effectively eliminates the phenomenon of delayed prediction, can capture transient dynamic characteristics in real time, improves the accuracy and robustness of prediction, and adapts to high-frequency disturbances and sudden changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070183A_ABST
    Figure CN123070183A_ABST
Patent Text Reader

Abstract

This invention discloses a multivariate time series forecasting method based on a dual-branch liquid state-space model. It employs an explicit variable decoupling strategy, synchronously segmenting the input multivariate time series data into parallel autoregressive and multivariate residual branches, achieving a physical separation between the pure historical evolution trend and external nonlinear forcing features. An input-driven second-order polynomial upgrade operator is introduced to model the nonlinear convection term of local hidden features, generating a Poly-B liquid permeation valve that dynamically deforms in real time with the input intensity. This permeation valve is then superimposed on the global static convolution output of the state space, achieving dynamic modulation of the microscopic convolution kernel weights. Finally, an adaptive neural gating mechanism at the terminal state dynamically fuses the baseline autoregressive prediction value and the externally forced residual prediction value. This invention can be widely applied to complex dynamic systems such as meteorological dynamics simulation, urban traffic network flow control, and communication network congestion prediction, significantly improving prediction accuracy and system robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence deep learning and multivariate time series prediction technology. Specifically, it relates to a multivariate time series prediction method based on a two-branch liquid state space model. Through macroscopic two-branch explicit variable decoupling design and microscopic second-order polynomial liquid permeation valve regulation, it can widely support multivariate time series analysis tasks such as traffic flow velocity control in urban traffic networks, evolution prediction of meteorological dynamic indicators, tracking of epidemiological transmission trends, and dynamic traffic prediction of congestion at routing nodes in modern communication backbone networks. It has achieved prediction accuracy and generalization robustness that significantly surpasses existing advanced baseline models. Background Technology

[0002] In complex multidimensional nonlinear dynamic systems, multivariate time series forecasting (MTSF) forms the cornerstone of critical decision-making and proactive management. For example, in meteorological thermodynamics analysis, accurate time series forecasting enables early monitoring and prevention of severe weather by jointly analyzing complex correlation indicators such as temperature, humidity, and air pressure. In epidemiological surveillance, precise evolutionary extrapolation of multidimensional demographic indicators such as the proportion of outpatients and the number of influenza-like illnesses can provide forward-looking guidance for the dynamic allocation of medical resources during sudden viral outbreaks. In modern network communication infrastructure, online short-term forecasting of high-dimensional source-destination (OD) traffic matrices between backbone routing nodes is crucial for avoiding severe network-wide congestion and dynamically allocating global bandwidth. With the rapid development of various high-frequency sensors and ubiquitous sensing networks, industry has accumulated massive amounts of non-stationary time series data with complex spatial topologies or physical couplings, laying a solid foundation for the application of deep learning models in the field of forecasting.

[0003] Currently, the mainstream deep learning architectures for multivariate time series prediction are recurrent neural networks (RNNs), temporal convolutional networks (TCNs), and Transformer networks based on self-attention mechanisms. In particular, the Transformer, with its powerful global dependency capture capability, has established a dominant position in long-term series prediction. In recent years, to balance the long-range memory capability and computational efficiency of high-frequency, massive time-series data, Structured State-Space Models (SSMs, such as the S4 model and the Mamba model) have emerged as a more efficient alternative. SSMs, through elegant discretization of continuous-time systems, utilize the Fast Fourier Transform (FFT) in the frequency domain to achieve an efficient convolutional flow with a time complexity of O(Llog L), greatly alleviating the computational bottleneck of traditional attention mechanisms, which increases quadratically with the backtracking window length.

[0004] However, despite the significant progress made by existing Transformer, MLP, or SSM-type time series models in traditional long-term stationary time series prediction tasks, these models collectively reveal two fatal theoretical flaws when deployed in real-world short-term prediction environments characterized by high chaos, high variability, and rich local mutations: First, the rigidity of representation at the microscopic level. Both the projection weight matrix of the Transformer and the core operators of the standard SSM model are inherently subject to the strict constraints of linear time-invariant (LTI) properties. The parameters of their convolutional kernels in the spatial or channel dimensions are globally static and cannot change over time. When dealing with ubiquitous transient strong disturbances, system transitions, or sporadic mutations (such as sudden congestion during holidays, sudden outbreaks of epidemics, or high-frequency, violent pulses in network traffic) in real-world chaotic systems, these rigid filters often forcibly smooth out these short-term transient dynamic characteristics, leading to severe instantaneous prediction errors. Second, the variable dependency dilemma at the macroscopic level. Existing multivariate prediction architectures (such as iTransformer and Channel-Dependent models) typically tend to force all feature variables to be mapped into the same tightly entangled shared network black box. Since neural networks naturally tend to fit smooth signals with low variance and high autocorrelation during optimization, this state of total variable entanglement causes the model to over-rely on safe, conservative autoregressive historical values, misjudging external high-frequency strong disturbances as useless noise and filtering them out. The direct visual and quantitative manifestation of this is the "lag prediction" phenomenon, where the model's output prediction curve is merely a lagging translation of historical ground values, failing to truly capture abrupt changes in future trends. Summary of the Invention

[0005] To address the technical shortcomings of existing deep learning time series prediction models, such as rigid microscopic representations and entangled macroscopic features, which lead to an inability to capture short-term mutations and severe lag prediction, this invention proposes a multivariate time series prediction method and system based on a bi-branch liquid state-space model. This method fundamentally breaks the traditional linear time-invariant (LTI) constraint, realizes a microscopic operator that senses dynamic input deformation and an explicitly decoupled macroscopic branch, and can effectively eliminate the lag effect in time series prediction.

[0006] The technical solution adopted in this invention is as follows:

[0007] A multivariate time series forecasting method based on a bibranch liquid state-space model includes the following processing steps: Step 1: Obtain batches of historical observation data X ∈ R for multivariate time series. B×L×DWhere B is the batch size, L is the backtracking window length, and D is the feature dimension of the input features; the multivariate time series prediction task is defined as learning a direct mapping function f from the historical observation data batch X to the target variable at the next T time steps, in order to predict the target variable prediction value Ŷ ∈ R at the next T time steps. B×T×1 Step 2: Employ an explicit variable decoupling strategy to perform macroscopic-level dual-branch decoupling processing on the input historical observation data batch X, synchronously segmenting it and inputting it into two parallel autoregressive branches and multivariate residual branches, thereby separating the overall dynamic physics of the system into pure historical evolution trends and nonlinear perturbations of external characteristics; Step 3: Within the autoregressive branch and the multivariate residual branch, extract dynamic time-varying features through independently weighted liquid core modules, and model the local hidden feature vectors using nonlinear convection terms by introducing an input-driven second-order polynomial upgrade operator, generating a Poly-B liquid permeation valve that dynamically deforms in real time with the input intensity, and superimposing the Poly-B liquid permeation valve onto the static baseline output of global convolution in the state space to obtain the dynamic modulation feature sequence corresponding to each branch. Step 4: Extract the terminal hidden states of the autoregressive branch and the multivariate residual branch at the last time step, respectively, and map them through a linear projection layer to generate the baseline autoregressive prediction value and the externally forced residual prediction value. Based on the final time step features of the input window, dynamically fuse the baseline autoregressive prediction value and the externally forced residual prediction value through an adaptive neural gating mechanism of terminal state to obtain the final target variable prediction value Ŷ. Step 5: Construct a composite loss function containing the main training objective and auxiliary monitoring indicators in the normalized latent space, calculate the change of physical gradient through the gradient backpropagation algorithm, and update all learnable parameters of each functional module and network layer in the method synchronously until the overall model training converges.

[0008] In step two, the macroscopic-level two-branch decoupling process specifically includes: in the autoregressive branch, extracting the univariate sequence of the target variable in channel c of the historical observation data batch X through variable slicing operations, and constructing the autoregressive branch input matrix X. (AR) = X[:, :, c] ∈ R B×L×1 This is used to isolate and capture the pure historical evolution inertia of the target variable; in the multivariate residual branch, the complete batch of multivariate historical observation data is directly used as the multivariate residual branch input matrix X. (MR) = X ∈ R B×L×D It is used to specifically capture nonlinear external forcing and multivariate residuals from all feature channels, thereby severing the feature entanglement of autoregressive inertia and external nonlinear noise in the same network space.

[0009] In step three, the specific micro-operations of the liquid core module in extracting dynamic time-varying features include: denoting the input sequence matrix of any branch as the generalized sequence input Z ∈ R. B×L×din First, the learnable input projection matrix W is used. in and bias vector b in By mapping Z to a unified high-dimensional hidden space through a linear layer, we obtain the high-dimensional hidden space matrix U = ZW. in + b in ∈ R B×L×dmodel , where d in Let d be the input feature dimension of the corresponding branch. model The high-dimensional hidden layer dimension is defined; the high-dimensional hidden space matrix U is sliced ​​along the time axis using time steps to extract the local hidden feature vector u(t) = U[:, t, :] ∈ R at any specified observation time t ∈ [1, L]. B×dmodel This serves as the input driving feature for the microscopic liquid transformation at that moment.

[0010] In step three, the nonlinear convection term modeling specifically includes: introducing a liquid polynomial order parameter k=2, and cascading and combining a first-order linear hidden feature with a second-order nonlinear time-transitional interaction term with a delay parameter τ to upgrade and map the local hidden feature vector u(t) to a higher-order nonlinear sensing space, thereby obtaining the upgraded polynomial hidden feature vector U. poly (t): U poly (t) = u(t) + α(u(t) ⊙ u(t-τ)) ∈ R B×dmodel In the formula, α is a learnable convection scaling factor used to control the oscillation intensity of nonlinear fluid interaction, τ is a historical time delay parameter, and ⊙ represents the element-level Hadamard product; wherein the second-order nonlinear trans-time interaction term u(t) ⊙ u(t-τ) corresponds to the nonlinear convection acceleration term in fluid mechanics, making the polynomial hidden feature vector U poly (t) can simultaneously capture the absolute numerical features of the hidden state and the local dynamic gradient that flows over time, thus stimulating the model to adapt to sudden changes in chaotic time series.

[0011] In step three, the process of generating and dynamically modulating the Poly-B liquid permeation valve that dynamically deforms in real time with the input intensity specifically includes: synchronously inputting the high-dimensional hidden space matrix U into the standard structured state space model S4 module for global convolution calculation, using frequency domain fast Fourier transform (FFT) to achieve long-range historical inertial capture with a computational complexity of O(L log L), and outputting the global static reference matrix O. S4∈ R B×L×dmodel Along the time axis, the global static reference matrix O S4 Slice the data to obtain the static output feature vector o at time t. S4 (t) = O S4 [:, t, :]; utilizing the upgraded polynomial hidden feature vector U poly (t) After the learnable parameter matrix W u The linear mapping is calculated using the hyperbolic tangent activation function tanh, generating a real-time dynamically adaptive characteristic adjustment control signal. This signal is used as the time-varying dynamic permeation counterpart of the input control matrix B in the standard state-space differential equation, defined as the Poly-B liquid permeation valve signal ΔB(t): ΔB(t) = tanh(U poly (t)W u ) ∈ R B×dmodel The Poly-B liquid permeation valve signal ΔB(t) is compared with the learnable fluid permeation matrix W. p The product is multiplied and applied to the static output feature vector o in a dot product and unit gain superposition manner. S4 On (t), the time-varying liquid feature vector o after local fluid adaptive modulation is obtained. polyb (t): o polyb (t) = o S4 (t) ⊙ (1 + ΔB(t)W p ) ∈ R B×dmodel After traversing and calculating the time series of all moments within the backtracking window, the time-varying liquid feature vectors o corresponding to each moment are... polyb (t) By sequentially splicing along the time axis, the terminal branch output matrix O of the liquid core module is reconstructed. polyb ∈ R B ×L×dmodel This enables input-aware deformation of micro-convolutional kernel weights.

[0012] In step four, the process of generating predicted values ​​through mapping specifically includes: from the output matrix O of the liquid core module within the autoregressive branch... polyb (AR) In the process, the feature vector at the last time step L is extracted as the autoregressive terminal hidden state H. last (AR) = O polyb (AR) [:, L, :] ∈ R B×dmodel Using the corresponding learnable autoregressive prediction weight matrix W AR With bias term b ARPerform linear mapping projection to generate baseline autoregressive predicted values ​​Ŷ AR = H last (AR) W AR + b AR ∈R B×T×1 ; Synchronously from the output matrix O of the liquid core module inside the multivariate residual branch polyb (MR) In the process, the feature vector at the last time step L is extracted as the hidden state H of the multivariate residual terminal. last (MR) = O polyb (MR) [:, L, :] ∈ R B ×dmodel Using the corresponding learnable multivariate residual prediction weight matrix W MR With bias term b MR Perform linear mapping projection to generate predicted externally forced residual values. MR = H last (MR) W MR + b MR ∈ R B×T×1 .

[0013] In step four, the process of dynamic fusion of the neural gating mechanism specifically includes: extracting all channel feature vectors x at the final time L in the original input historical observation data batch X. L = X[:, L, :] ∈ R B×D As an external mechanism for perceiving the current system's environment; the feature vector x L The input is fed into the terminal state adaptive multilayer perceptron, and firstly, the weight matrix W of the first hidden layer is used. h and bias b h A linear mapping is performed, followed by nonlinear activation using the modified linear unit ReLU, and then the second mapping layer weight matrix W is applied. g and bias b g The transformation is performed, and finally, the Sigmoid activation function σ is used for amplitude limiting and compression, thereby generating a dynamic gating coefficient g whose value range is strictly limited to the interval (0, 1): g = σ(W g ReLU(W h x L + b h ) + b g Using the dynamic gating coefficient g, the baseline autoregressive predicted value Ŷ is calculated. AR Compared with the predicted value of externally forced residualsMR Dynamic matching and fusion are performed, and the final fusion output can effectively eliminate the time lag phenomenon in the target sequence prediction result Ŷ: Ŷ = g ⊙ Ŷ AR + (1 - g) ⊙(Ŷ AR + Ŷ MR ).

[0014] In step five, the process of constructing the composite loss function specifically includes: using the mean absolute error (MAE) as the main training loss function L for global collaborative optimization and updating of model parameters. MAE Its mathematical expression is: L MAE = [1 / (B· T)] Σ b=1 B Σ t=1 T |ŷ b,t - y b,t In the formula, ŷ b,t and y b,t These represent the predicted value of the target variable and the actual ground value of the b-th data batch at the t-th prediction time step, respectively, under the normalized latent spatial scale. The mean squared error (MSE) is introduced simultaneously as an auxiliary inspection and monitoring index for the model, used to impose a strong, quadratic-level structured penalty on extreme anomalous jumps, sudden peak mutations, or outlier drastic fluctuations in non-stationary dynamic systems, preventing the model from falling into local overfitting. The numerical calculation of the main training loss function and the auxiliary inspection and monitoring index, as well as the corresponding error gradient backpropagation, are all performed after being reconstructed to the actual physical scale through rigorous channel-level inverse normalization processing to ensure that the calculated optimization gradient has clear industrial interpretability.

[0015] The beneficial effects of this invention are as follows: 1. Macroscopic explicit variable decoupling: This invention breaks the black box of traditional multivariate prediction network training with all variables. Through physical partitioning, the evolution history of the target and external multivariate disturbances are completely separated in an independent parallel operator space without shared weights. This greatly eliminates the inherent defect of deep networks that actively sacrifice and smooth out high-frequency, high-value disturbance information due to greedy fitting of low-variance trends, fundamentally eradicating the "lagging prediction" problem. 2. Microscopic fluid liquid state adaptation: Through cascaded second-order polynomial nonlinear interactive convection terms, the model of this invention can calculate and sense the flow gradient of the time-hidden feature matrix in real time. The resulting Poly-B liquid permeation valve allows the convolution kernel of the originally time-static structured state-space model to deform in real time during operation with the intensity of the input signal oscillation. When encountering sudden abnormal pulses, the network can actively adjust its liquid damping state, rapidly adjust its effective fluid time constant, and prioritize the capture of recent extremely high-frequency features. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] like Figure 1 As shown, the multivariate time series prediction method based on a bi-branch liquid state-space model proposed in this invention demonstrates excellent system architecture completeness and engineering deployment feasibility in practical industrial applications. This section will provide a comprehensive technical overview of the engineering implementation details, including system environment configuration, data stream preprocessing, core execution logic, and the smooth switching architecture between offline training and online streaming inference, without delving into the specific experimental data, in order to highlight the pure technical architecture of this invention.

[0019] I. System Hardware and Software Operating Environment and Deployment Specifications To ensure the high-throughput, low-latency, and stable operation of the prediction method of this invention in actual industrial production environments (such as substation automation status monitoring systems, intelligent transportation central control and dispatch platforms, and large-scale cloud data center traffic firewalls), the system is recommended to be deployed in the following computing environment: In terms of hardware, a multi-core industrial-grade CPU with 16 or more cores is recommended for the central processing unit (CPU) to concurrently execute streaming data reading, time-sliding window slicing preprocessing, and asynchronous distribution of dual-branch data. Simultaneously, the system is equipped with a high-performance graphics processing unit (GPU) or a dedicated AI acceleration chip (such as a tensor processor chip) to support the continuous discretization of the large-scale state space matrices A, B, and C in the liquid core module, the generation of high-order polynomial fluid convolution kernels, and the hardware acceleration of high-frequency fast Fourier transform (FFT), ensuring that single-step inference latency is strictly suppressed to within single-digit milliseconds. In terms of software environment, the underlying layer adopts an industrial-grade operating system with high real-time performance assurance, and the middle layer deep learning framework adopts a high-performance tensor computing platform that supports high-order differential calculation and automatic differentiation of continuous implicit operators. It is combined with the underlying matrix acceleration library (such as NVIDIA CUDA and cuDNN acceleration kit) to achieve integrated hardware and software collaborative and efficient computing.

[0020] II. Network Layer Parameter Initialization and Multidimensional Sliding Window Preprocessing Mechanism When the system is loaded and enters the runtime, the initialization program is first called to assign initial values ​​to the parameters of each component within the model. Since the micro-operators of this invention rely on the long-range memory capability of the Structured State-Space Model (SSM), the SSM state transition matrix A within the autoregressive branch and the multivariate residual branch is initialized as a HiPPO matrix (High-order Polynomial Projection Operator) with strictly mathematical orthogonality and optimal polynomial projection properties. This special matrix initialization mechanism ensures that the hidden layer states do not experience gradient explosion or gradient annihilation phenomena common in traditional recurrent neural networks during continuous-time integration. For the weight matrices of other ordinary linear projection layers and multilayer perceptron gated layers, a Gaussian or uniform scaling method is uniformly used for random initialization, and the learnable bias vector b is initially cleared to zero, thus ensuring a stable signal flow in the early stages of network evolution. In the data input preprocessing stage, the system receives the raw multidimensional temporal feature stream from an external sensor array or distributed bus. The system internally hosts a dynamic temporal sliding window slicer based on a circular buffer, which continuously slides and truncates data along the time axis at fixed strides, thereby generating continuous batches of data tensors X∈R. B×L×D The extracted historical time series segments undergo full-channel mean-variance standard normalization, which strictly constrains the data amplitude within a standardized latent scale that is easily absorbed and updated by deep neural networks, effectively eliminating numerical bias interference between multidimensional industrial indicators caused by different physical units of measurement.

[0021] III. Full Flow and Topology Evolution Logic of the Dual-Branch Core Algorithm The specific execution flow of this invention is described below: Step 1: Data Acquisition and Problem Definition. The time-series data acquisition module reads the current standardized multivariate historical time series batch X ∈ R from the aforementioned sliding window slicer. B×L×D The feature dimension D exhibits extremely high generalization adaptability depending on the target domain, and can cover any multivariate dynamic system, including but not limited to weighted and unweighted epidemiological ratios, multidimensional meteorological thermodynamic dynamic indicators, and high-dimensional routing traffic of the backbone network. The mapping objective of this invention is to efficiently calculate and output the target prediction value Ŷ ∈ R for future time steps. B×T×1 Step Two: Macroscopic Two-Branch Decoupling. The macroscopic two-branch decoupling module performs explicit variable slicing on the input tensor X. In the autoregressive branch, the model extracts the univariate time series of the target independent variable through single-channel indexing, constructing matrix X. (AR) = X[:, :, c] ∈ R B×L×1This ensures the purity of historical inertia; in the multivariate residual branch, the complete full-channel multidimensional features are directly fed as input X. (MR) = X ∈ R B×L×D It is used to capture the multivariate residuals of multidimensional joint nonlinear perturbations of external variables.

[0022] Step 3: Time-varying modulation of the microscopic liquid operator. After entering the two branches, each branch is equipped with a set of structurally isomorphic but completely decoupled liquid core modules. The generalized sequence Z ∈ R is used as the input sequence of any branch. B×L×din For example: First, using a fully connected hidden projection layer, Z is mapped to a unified high-order hidden representation matrix U = ZW. in +b in ∈ R B×L×dmodel Extract the local vector u(t) = U[:, t, :] ∈ R at observation time t along the time axis. B ×dmodel The input-driven second-order polynomial upgrading operator then performs a nonlinear upgrade on it: U poly (t) = u(t) + α(u(t) ⊙ u(t-tau)). The nonlinear second-order term simulates the convection term in fluid dynamics, causing the hidden layer to drastically amplify the hidden flow gradient between adjacent time points. Simultaneously, the complete hidden space matrix U is input into the standard structured state-space model S4 module, where a long-range temporal static global convolution is performed in the frequency domain to extract the output o. S4 (t) =O S4 [:, t, :]. Subsequently, the polynomial eigenvector U poly (t) The Poly-B liquid permeation valve signal ΔB(t) is generated by projection and varies online with the input intensity: ΔB(t) = tanh(U poly (t)W u Finally, by controlling the valve at the element-level dot product level, dynamic time-varying modulation is applied to the static convolution output: o polyb (t) = o S4 (t) ⊙ (1 + ΔB(t)W p After completing all time steps within the backtracking window L, the results are sequentially concatenated and reconstructed into the terminal branch output matrix O. polyb ∈ R B×L×dmodel .

[0023] Step 4: Terminal Gated Fusion Prediction. The terminal gated fusion prediction module extracts the vectors of the output matrices corresponding to the two branches at the last backtracking time step L, as their respective representations of the terminal's hidden state H after long-range evolution. last (AR) = O polyb (AR) [:, L, :] and H last (MR) = O polyb (MR) [:, L, :]. Then, through their respective predictive projection linear layers, the baseline autoregressive predicted value Ŷ_AR = H is output. last (AR) W_AR + b_AR and the predicted value of the externally forced residual Ŷ_MR = H last (MR) W_MR + b_MR. To dynamically adjust the baseline-to-residual ratio during fast mechanism transitions in a non-stationary system, the system is designed based on the true state x at the final time step of the current input window. L A multilayer perceptron gating network of = X[:, L, :]: g = σ(W g ReLU(W h x L + b h ) + b g Through external mechanism perception, the gating coefficient g within the (0, 1) interval is calculated. The final fusion formula is: Ŷ = g ⊙ Ŷ_AR + (1 - g) ⊙ (Ŷ_AR + Ŷ_MR). When the external environment is extremely stable, g approaches 1, and the model reverts to conservative autoregressive evolution; however, once faced with sudden large-scale fluctuations, the gating coefficient g automatically and instantaneously shrinks, strongly injecting the externally forced multivariate residual Ŷ_MR, making the predicted response curve completely synchronized with the actual mutation, eliminating the lag prediction phenomenon. Step 5: Composite objective function optimization and update. During the model training phase, the multi-objective collaborative optimization and update module constructs a composite loss function under the normalized latent scale. The main training loss function adopts the mean absolute error (MAE). The mean squared error (MSE) is introduced synchronously as an auxiliary penalty term, which applies a structured heavy penalty to the outlier extreme prediction errors generated during mutations through the squared term property, preventing gradient saturation and local overfitting. All optimization updates are performed using the backpropagation algorithm and the AdamW optimizer.

[0024] IV. Smooth Switching Architecture Between Offline Training and Online Real-Time Inference To ensure the method of this invention truly aligns with modern industrial big data production deployments, the system is designed as a flexible operating architecture capable of seamless switching between two stages: In the offline training stage, the system relies on a high-computing-power historical database to transmit historical data in parallel using batch and large tensor batch formats. During this time, the multi-objective collaborative optimization update module operates at full capacity, calculating the main loss L. MAE and auxiliary loss L MSE Using the backpropagation mechanism, all learnable weight matrices (W) within the system are processed. in W u W p W AR W MR W h W g The system performs rapid gradient descent iterations on the weight constants, bias matrices, and state transition parameters of the S4 operator (as well as the corresponding bias vectors). Once the network training meets the set convergence criteria, the system automatically activates the static sequence storage engine, persistently writing all converged weight constants, bias matrices, and S4 operator state transition parameters into a local lightweight model dictionary file (.pth or .bin format), completing the hard-coding and solidification of knowledge. During the online real-time inference stage, the system completely shuts down the central multi-objective optimization update module (step five) and places it in a dormant state. The online high-frequency inference engine only needs to load the offline solidified ultra-lightweight model dictionary file. When external production buses or real-time sensor pipelines pump in data with high-frequency, single-step streaming media, the sliding window preprocessing mechanism immediately extracts historical segments within the latest time span, routes them in a very short unidirectional forward propagation network, and executes them concurrently in GPU / CPU accelerated operators within seconds, instantly outputting the current moment's lag-free look-ahead prediction value for the future. This dual-stage flexible switching mechanism, which involves deep offline knowledge accumulation and high-frequency, seamless online reasoning, endows the invention with extremely high robustness, industrial applicability, and a strong all-scenario survival barrier.

Claims

1. A multivariate time series prediction method based on a two-branch liquid state-space model, characterized in that, Includes the following steps: Step 1: Obtain batches of multivariate time series historical observation data X ∈ R B×L×D Where B is the batch size, L is the backtracking window length, and D is the feature dimension of the input feature; The multivariate time series forecasting task is defined as learning a direct mapping function f from the historical observation data batch X to the target variable at the next T time steps, in order to predict the target variable Ŷ ∈ R at the next T time steps. B×T×1 ; Step 2: The input historical observation data batch X is subjected to macroscopic-level dual-branch decoupling processing using an explicit variable decoupling strategy. It is synchronously split and input into two parallel autoregressive branches and multivariate residual branches, thereby separating the overall dynamics of the system into pure historical evolution trend and nonlinear perturbation of external characteristics. Step 3: Within the autoregressive branch and the multivariate residual branch, dynamic time-varying features are extracted using independently weighted liquid core modules. A second-order polynomial upgrade operator driven by the input is introduced to model the local hidden feature vectors using nonlinear convection terms, generating a Poly-B liquid permeation valve that dynamically deforms in real time with the input intensity. The Poly-B liquid permeation valve is then superimposed on the static baseline output of the global convolution in the state space to obtain the dynamic modulation feature sequence corresponding to each branch. Step 4: Extract the terminal hidden states of the autoregressive branch and the multivariate residual branch at the last time step, respectively, and generate the baseline autoregressive prediction value and the externally forced residual prediction value through a linear projection layer. Based on the final time step features of the input window, the baseline autoregressive prediction value and the externally forced residual prediction value are dynamically fused through the terminal state adaptive neural gating mechanism to obtain the final target variable prediction value Ŷ. Step 5: Construct a composite loss function containing the main training objective and auxiliary monitoring indicators in the normalized latent space, and calculate the change of physical gradient through the gradient backpropagation algorithm. In this way, update all learnable parameters of each functional module and network layer in the method until the overall model training converges.

2. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 1, characterized in that, In step two, the macroscopic-level two-branch decoupling process specifically includes: In the autoregressive branch, the univariate sequence of the target variable in channel c of the historical observation data batch X is extracted through variable slicing to construct the autoregressive branch input matrix X. (AR) = X[:, :, c] ∈ R B×L×1 This is used to isolate and capture the historical evolutionary inertia of the pure target variable; In the aforementioned multivariate residual branch, the complete batch of multivariate historical observation data is directly used as the input matrix X of the multivariate residual branch. (MR) = X ∈ R B×L×D It is used to specifically capture nonlinear external forcing and multivariate residuals from all feature channels, thereby severing the feature entanglement of autoregressive inertia and external nonlinear noise in the same network space.

3. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 1, characterized in that, In step three, the specific microscopic operations for the liquid core module to extract dynamic time-varying features include: Let the input sequence matrix of any branch be denoted as the generalized sequence input Z ∈ R. B×L×din First, the learnable input projection matrix W is used. in and bias vector b in By mapping Z to a unified high-dimensional hidden space through a linear layer, we obtain the high-dimensional hidden space matrix U = ZW. in + b in ∈ R B×L×dmodel , where d in Let d be the input feature dimension of the corresponding branch. model For the higher-dimensional hidden layer dimension; The high-dimensional hidden space matrix U is sliced ​​along the time axis to extract the local hidden feature vector u(t) = U[:, t, :] ∈ R at any specified observation time t ∈ [1, L]. B×dmodel This serves as the input driving feature for the microscopic liquid transformation at that moment.

4. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 3, characterized in that, In step three, the nonlinear convection term modeling specifically includes: By introducing a liquid polynomial order parameter k=2, and cascading a first-order linear hidden feature with a second-order nonlinear time-transitional interaction term with a delay parameter τ, the local hidden feature vector u(t) is upgraded and mapped to a higher-order nonlinear sensing space, resulting in the upgraded polynomial hidden feature vector U. poly (t): U poly (t) = u(t) + α(u(t) ⊙ u(t-τ)) ∈ R B×dmodel In the formula, α is a learnable convection scaling factor used to control the oscillation intensity of nonlinear fluid interaction, τ is a historical time delay parameter, and ⊙ represents the element-level Hadamard product; wherein the second-order nonlinear trans-time interaction term u(t) ⊙ u(t-τ) corresponds to the nonlinear convection acceleration term in fluid mechanics, making the polynomial hidden feature vector U poly (t) can simultaneously capture the absolute numerical features of the hidden state and the local dynamic gradient that flows over time, thus stimulating the model to adapt to sudden changes in chaotic time series.

5. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 4, characterized in that, In step three, the process of generating and dynamically modulating the Poly-B liquid permeation valve that dynamically deforms in real time according to the input intensity specifically includes: The high-dimensional hidden space matrix U is synchronously input into the standard structured state space model S4 module for global convolution calculation. The Fast Fourier Transform (FFT) in the frequency domain is used to achieve long-range historical inertial capture with low computational complexity, outputting a global static reference matrix O. S4 ∈ R B×L×dmodel ; Along the time axis, the global static reference matrix O S4 Slice the data to obtain the static output feature vector o at time t. S4 (t) = O S4 [:, t, :]; Using the upgraded polynomial hidden feature vector U poly (t) After the learnable parameter matrix W u The linear mapping is calculated using the hyperbolic tangent activation function tanh, generating a real-time dynamically adaptive characteristic adjustment control signal. This signal is used as the time-varying dynamic permeation counterpart of the input control matrix B in the standard state-space differential equation, defined as the Poly-B liquid permeation valve signal ΔB(t): ΔB(t) = tanh(U poly (t)W u ) ∈ R B×dmodel The Poly-B liquid permeation valve signal ΔB(t) is compared with the learnable fluid permeation matrix W. p The product is multiplied and applied to the static output feature vector o in a manner that combines dot product and superposition of unit gain. S4 On (t), the time-varying liquid feature vector o after local fluid adaptive modulation is obtained. polyb (t): o polyb (t) = o S4 (t) ⊙ (1 + ΔB(t)W p ) ∈ R B×dmodel After traversing and calculating the time series of all moments within the backtracking window, the time-varying liquid feature vectors o corresponding to each moment are... polyb (t) By sequentially splicing along the time axis, the terminal branch output matrix O of the liquid core module is reconstructed. polyb ∈ R B×L×dmodel This enables input-aware deformation of micro-convolutional kernel weights.

6. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 1, characterized in that, In step four, the process of generating predicted values ​​through mapping specifically includes: The output matrix O of the liquid core module within the autoregressive branch polyb (AR) In the process, the feature vector at the last time step L is extracted as the autoregressive terminal hidden state H. last (AR) = O polyb (AR) [:, L, :] ∈ R B×dmodel Using the corresponding learnable autoregressive prediction weight matrix W AR With bias term b AR Perform linear mapping projection to generate baseline autoregressive predicted values ​​Ŷ AR = H last (AR) W AR + b AR ∈ R B×T×1 ; Synchronous output matrix O from the liquid core module within the multivariate residual branch polyb (MR) In the process, the feature vector at the last time step L is extracted as the hidden state H of the multivariate residual terminal. last (MR) = O polyb (MR) [:, L, :] ∈ R B ×dmodel Using the corresponding learnable multivariate residual prediction weight matrix W MR With bias term b MR Perform linear mapping projection to generate predicted externally forced residual values. MR = H last (MR) W MR + b MR ∈ R B×T×1 .

7. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 6, characterized in that, In step four, the process of dynamic fusion of the neural gating mechanism specifically includes: Extract all channel feature vectors x at the final time L from the original input historical observation data batch X. L =X[:, L, :] ∈ R B×D As an external mechanism for perceiving the current system's environment; The feature vector x L The input is fed into the terminal state adaptive multilayer perceptron, and firstly, the weight matrix W of the first hidden layer is used. h and bias b h A linear mapping is performed, followed by nonlinear activation using the modified linear unit ReLU, and then the weight matrix W of the second mapping layer is applied. g and bias b g The transformation is performed, and finally, the Sigmoid activation function σ is used for amplitude limiting and compression, thereby generating a gating coefficient g whose value range is strictly limited to the interval (0, 1): g = σ(W g ReLU(W h x L + b h ) + b g ) The baseline autoregressive predicted value Ŷ is predicted using the dynamic gating coefficient g. AR Compared with externally forced residual predictions MR Dynamic matching and fusion are performed, and the final fusion output is a target sequence prediction result that can eliminate the time lag phenomenon. Ŷ = g ⊙ Ŷ AR + (1 - g) ⊙ (Ŷ AR + Ŷ MR )。 8. The multivariate time series prediction method based on a bi-branch liquid state-space model according to claim 1, characterized in that, In step five, the process of constructing the composite loss function specifically includes: The mean absolute error (MAE) is used as the main training loss function L for global co-optimization and updating of model parameters. MAE Its mathematical expression is: L MAE = [1 / (B · T)] S b=1 B S t=1 T |ŷ b,t -y b,t | In the formula, ŷ b,t and y b,t and represent the predicted and actual ground values ​​of the target variable for the b-th data batch at the t-th prediction time step in the future, respectively, under the normalized latent scale; The mean squared error (MSE) is introduced as an auxiliary monitoring metric for the model. It is used to apply a strong, structured penalty of the quadratic level to extreme abnormal jumps, sudden peak mutations, or outlier violent fluctuations in non-stationary dynamic systems, so as to prevent the model from falling into local overfitting. The numerical calculation of the main training loss function and the auxiliary monitoring index, as well as the corresponding error gradient backpropagation, are all performed after being reconstructed to the actual physical scale through strict channel-level inverse normalization processing, to ensure that the calculated optimization gradient has clear industrial interpretability.

9. A multivariate time series prediction system based on a bi-branch liquid state-space model that implements the method described in any one of claims 1-8, characterized in that, include: The time series data acquisition module is used to acquire batches of multivariate time series historical observation data X ∈ R. B×L×D ; The macroscopic dual-branch decoupling module is used to split the historical observation data batch X into a univariate sequence matrix X of pure objective variables using an explicit variable decoupling strategy. (AR) With a multivariate feature matrix X containing complete input features (MR) And simultaneously fed into the autoregressive branch and the multivariate residual branch, which operate independently of each other; The local fluid adaptive feature extraction module, built within the autoregressive and multivariate residual branches, includes a liquid core module. This module transforms the input static time-invariant state space convolution kernel into a time-varying liquid adaptive feature representation that dynamically deforms with the input data intensity through linear hidden space projection, time-step piecewise extraction, second-order polynomial feature upgrading, and real-time calculation of the Poly-B liquid permeation valve. It outputs the corresponding dynamically modulated feature sequences O. polyb (AR) and O polyb (MR) ; The end-gated fusion prediction module is used to extract the terminal hidden state of the output matrix corresponding to each of the two branches at the final time step and linearly map it to generate the baseline autoregressive prediction value. AR Compared with externally forced residual predictions MR Meanwhile, the gating coefficient g is calculated and generated using a multilayer perceptron based on the final time state perception, and the final target variable prediction value Ŷ is output through nonlinear gating weighted fusion. The multi-objective collaborative optimization and update module is used to construct a composite objective loss function with mean absolute error (MAE) as the main training loss and mean squared error (MSE) as an auxiliary indicator for abnormal jump penalty, and to perform collaborative updates until convergence.