Rapid time-space prediction method for power grid load based on TSMIX algorithm

Abstract

The invention discloses a rapid time-space prediction method for a power grid load based on a TSMIX algorithm, relates to the technical field of power grid load prediction, and is used for solving the problems that parallel resources cannot be fully utilized, the parameter scale is large, and the calculated amount is large in an existing traditional short-term load prediction method. Comprising the following steps: S1, processing original multivariable time sequence input data from an integrated energy system through an embedded layer, and mapping the original input data to a potential space with higher dimensionality; s2, capturing a sequence dependency relationship in each variable by adopting a time multi-layer perceptron (MLP); and S3, the output of the time multi-layer perceptron MLP is transmitted to the space multi-layer perceptron MLP, and the space multi-layer perceptron MLP carries out modeling on correlation between variables among different subsystems. And S4, decoding the output of the spatial multilayer perceptron MLP and generating prediction results of all target variables at the same time. According to the method, the calculation time is remarkably shortened under the condition that the precision is not reduced.

Description

Technical Field

[0001] This invention relates to the field of power grid load forecasting technology, specifically a fast spatiotemporal forecasting method for power grid load based on the TSMixer algorithm. Background Technology

[0002] Integrated energy systems (IES) have become an effective model for coordinating electricity, heating, and cooling networks to achieve efficient and low-carbon energy utilization. In recent years, urban IES have made significant progress through digital management platforms, intelligent dispatching, and data-driven optimization. In contrast, rural microgrids, a key component of the distributed energy landscape, are still in their early stages of development, facing challenges such as limited regulation capacity, large fluctuations in renewable energy, and insufficient measurement and decision support capabilities. As rural areas increasingly adopt photovoltaic, wind, and biomass energy resources to promote local energy self-sufficiency, accurate short-term forecasting is crucial for improving renewable energy utilization and maintaining stable microgrid operation. Therefore, developing rapid and efficient forecasting methods that can be embedded in resource-constrained rural IES has significant practical implications.

[0003] Short-term load forecasting plays a fundamental role in ensuring the economical dispatch and optimized operation of integrated energy systems, as it directly impacts the system's reliability and efficiency. However, the load of integrated energy systems is influenced by various factors, including climate conditions, seasonal variations, and user behavior, resulting in large load fluctuations and multiple types of coupling characteristics. These factors place stringent demands on forecasting methods, which must be able to provide accurate forecasts in a timely manner to support real-time dispatch decisions.

[0004] Traditional time series forecasting methods, such as the Autoregressive Integral Moving Average (ARIMA) model and Support Vector Machine (SVM), have limited ability to capture nonlinear and complex spatiotemporal relationships. In recent years, deep learning methods, especially Long Short-Term Memory (LSTM) networks and Transformer models, have demonstrated powerful feature extraction capabilities in load forecasting. However, LSTM networks rely on sequential computation, which limits their speed and prevents full utilization of parallel computing resources. While Transformer models can perform parallel computation, their large parameter size and high computational cost make them unsuitable for real-time forecasting in multivariate, highly coupled integrated energy system scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a fast spatiotemporal forecasting method for power grid load based on the TSMixer algorithm, which solves the problems of existing traditional short-term load forecasting methods that cannot fully utilize parallel resources and have large parameter scales, resulting in large computational loads.

[0006] The technical solution adopted by this invention to solve its technical problem is: a fast spatiotemporal prediction method for power grid load based on the TSMixer algorithm, comprising the following steps:

[0007] S1. The original multivariate time series input data from the integrated energy system is processed through the embedding layer, and the original input data is mapped to a higher-dimensional potential space.

[0008] S2. A time-multilayer perceptron (MLP) is used to capture the sequence dependencies within each variable.

[0009] S3. The output of the temporal multilayer perceptron (MLP) is passed to the spatial multilayer perceptron (MLP), which models the correlation between variables in different subsystems.

[0010] S4. The output of the decoding spatial multilayer perceptron (MLP) simultaneously generates the prediction results for all target variables.

[0011] Furthermore, the multivariate time series X is mapped to the embedding space through linear projection: (1); where, (2), B represents the data sample size; C represents the number of variables, i.e., the load type; L represents the length of the time series; R represents the data domain; and These are learnable parameters.

[0012] Furthermore, for Two fully connected layers (FC) with ReLU activation functions are applied: (3); where L represents the length of the time series; , W1 （t） and W2 （t） It is a weight matrix; and It is a deviation term; It is the activation function ReLU; It is the hidden dimension of temporal features; the output of the temporal multilayer perceptron (MLP) is: (4).

[0013] Furthermore, for a time step of variable vector Processing: (5); among which, , W1 （S） and W2 （S） It is the weight matrix s; and It is a deviation term; It is the dimension of hidden spatial features; the output of the spatial multilayer perceptron (MLP) is: (6).

[0014] Furthermore, the features processed by the Spatial Multilayer Perceptron (MLP) are mapped to the prediction output through a decoding layer: (7) The final multivariate time series prediction is generated through decoding. W （d） and These are learnable parameters; This indicates the length of the output sequence.

[0015] Furthermore, parallel computing is introduced for optimization in both temporal and spatial multilayer perceptron (MLP): For temporal MLP, multivariate time series data are processed... Adjusted to: (8); All variables are processed in parallel using the same multilayer perceptron (MLP): (9); Re-adjust the output as follows: (10).

[0016] The beneficial effects of this invention are: (1) Constructing a lightweight spatiotemporal architecture based on TSMixer: A lightweight prediction framework is designed for IES conformal features, in which the temporal pattern within each variable is modeled by an intra-variable multilayer perceptron (MLP), and the spatial dependency between variables is captured by an inter-variable multilayer perceptron (MLP). While keeping the number of parameters small, spatiotemporal features are effectively extracted. (2) Parallel acceleration strategy for real-time prediction: A parallel computing optimization scheme is proposed to break through the speed bottleneck of the serial structure and significantly reduce the computation time without reducing the accuracy. (3) Balancing high performance and low computational cost: This invention is superior to CNN, LSTM and Transformer in accuracy, while having higher computational efficiency and smaller model size. Attached Figure Description

[0017] Figure 1 This is a coupling diagram of the integrated energy subsystem;

[0018] Figure 2 This is an overall framework diagram of the method model of the present invention;

[0019] Figure 3 The graph shows the overall prediction performance based on the ASU-IES dataset;

[0020] Figure 4 Figure showing ablation study results based on the ASU-IES dataset;

[0021] Figure 5 This is a comparison chart of prediction times for different models. Detailed Implementation

[0022] The model of the method of this invention is developed based on the operating logic of the Integrated Energy System (IES), and its schematic diagram is shown below. Figure 1 As shown, the integrated energy system comprises electricity, gas, heating, and cooling subsystems, interconnected through various conversion and storage pathways. Energy from the electricity and gas networks is transported to conversion units such as electricity-to-cooling (P2C), electricity-to-heat (P2H), gas-to-electricity (G2P), gas-to-cooling (G2C), and gas-to-heat (G2H), and further supported by cross-sectoral coupling and dedicated storage devices. Figure 1 The overall framework of the method of this invention is as follows: Figure 2 As shown in the figure. This design ensures that the inherent temporal patterns and spatial correlations of the integrated energy system can be effectively captured. The following describes a fast spatiotemporal forecasting method for grid load based on the TSMixer algorithm.

[0023] S1. The original multivariate time series input data from the integrated energy system is processed through the embedding layer, and the original input data is mapped to a higher-dimensional potential space.

[0024] Suppose the input multivariate time series is: (1). Where B represents the data sample size; C represents the number of variables, i.e., the load type; L represents the length of the time series; and R represents the data domain. The multivariate time series X is mapped to the embedding space through linear projection: (2). Among them, and These are learnable parameters.

[0025] S2. A time-multilayer perceptron (MLP) is used to capture the sequence dependencies within each variable.

[0026] A temporal multilayer perceptron (MLP) is used to model the temporal dependencies within each variable. For each Where L represents the length of the time series, two fully connected layers (FC) with ReLU activation functions are applied: (3). Among them, , W1 （t） and W2 （t） It is a weight matrix; and It is a deviation term; It is the activation function ReLU; This is the hidden dimension of the temporal feature. The output of the Temporal Multilayer Perceptron (MLP) is: (4).

[0027] S3. The output of the temporal multilayer perceptron (MLP) is passed to the spatial multilayer perceptron (MLP), which models the correlation between variables in different subsystems.

[0028] Spatial multilayer perceptrons (MLPs) model the spatial dependencies between variables at each time step. For each time step... The corresponding variable vector The processing method is as follows: (5). Among them, , W1 （S） and W2 （S） It is the weight matrix s. as well as It is a deviation term; It is the dimension of hidden spatial features. The output of the spatial multilayer perceptron (MLP) is: (6).

[0029] S4. The output of the decoding spatial multilayer perceptron (MLP) simultaneously generates the prediction results for all target variables.

[0030] The decoding layer maps the features processed by the spatial multilayer perceptron (MLP) to the predicted output: (7). Among them, W （d） and These are learnable parameters; This indicates the length of the output sequence. The final multivariate time series prediction is generated through decoding.

[0031] To overcome the speed bottleneck of sequential computation, parallel computing is introduced for optimization in temporal multilayer perceptron (MLP) and spatial multilayer perceptron (MLP). For temporal MLP, multivariate time series... Instead of processing each variable independently, the process has been restructured as follows: (8). Allow all variables to be processed in parallel using the same multilayer perceptron (MLP): (9). Reorder the output as follows: (10). This method processes all variables in a batch computation, which theoretically can improve speed by C times compared to sequential processing.

[0032] The method of the present invention will be verified below. The verification objects include dataset description, parameter configuration, benchmark model and prediction speed comparison.

[0033] 1) Dataset: The Arizona State University Integrated Energy Systems (ASU-IES) dataset was used, which covers the period from January 1, 2019 to December 31, 2020, and contains a total of 17,544 hourly records. The data includes three main variables: electrical load, heat load, and cooling load.

[0034] 2) Parameter Configuration: To ensure fair comparison, all neural network-based methods used the same settings. The input sequence length was fixed at 96, and the prediction range was 96. The batch size was set to 32, and the dropout rate was 0.1. The optimizer used was Adam, and the loss function was mean squared error (MSE). An early stopping strategy was adopted, with a maximum of 10 training epochs and adaptive learning rate decay. All experiments were repeated three times to ensure stability. The experiments were implemented in PyTorch and executed on a single NVIDIA RTX 3060 GPU.

[0035] Model evaluation metrics include mean squared error (MSE) and mean absolute error (MAE): (11). (12). Among them, Represents the true value. This represents the predicted value.

[0036] 3) Benchmark Model: The proposed method is compared with the following state-of-the-art models.

[0037] BiLSTM A bidirectional long short-term memory network can simultaneously capture forward and backward dependencies of time series.

[0038] MICN: Multi-scale Interactive Convolutional Network, used to capture multi-scale temporal features.

[0039] Autoformer: A Transformer model based on sequence decomposition, suitable for long sequence prediction tasks.

[0040] Crossformer: A cross-dimensional Transformer variant for multivariate time series.

[0041] The experimental results and analysis of the validation method include overall performance and ablation studies.

[0042] 1) Overall performance: such as Figure 3 As shown, the mean squared error (MSE) and mean absolute error (MAE) results of each model on the ASU-IES dataset are presented. It can be seen that this invention achieves optimal results in both MSE and MAE, while requiring significantly fewer parameters and having a faster computation speed.

[0043] 2) Ablation Study: To verify the effectiveness of the temporal multilayer perceptron (MLP) and spatial multilayer perceptron (MLP) modules, this invention conducted ablation experiments. The results are as follows: Figure 4 As shown, the model performance significantly decreased when any module was removed, which confirms that the temporal modeling module can effectively capture the internal temporal patterns of a single variable, while the spatial modeling module can learn the dependencies between variables.

[0044] Furthermore, this invention also investigated the impact of module order. The original structure used a temporal multilayer perceptron (MLP) → spatial multilayer perceptron (MLP) order. When the order was changed to spatial multilayer perceptron (MLP) → temporal multilayer perceptron (MLP), performance decreased, indicating that performing temporal feature modeling first can better maintain the temporal consistency within variables.

[0045] 3) Prediction speed comparison: Compare the prediction speed of different models based on the same hardware and sample processing size. For example... Figure 5 As shown, the proposed model achieves the fastest prediction speed. This improvement is mainly due to the parallel computing strategy, which applies time modeling to the reshaped (B×C, L) sequence, thereby increasing the computation speed by a factor of C. In contrast, Autoformer and Crossformer are slower due to their large number of parameters, while BiLSTM is limited by its inherent serial computing mechanism.

[0046] This invention proposes a fast spatiotemporal forecasting method for power grid load based on the TSMixer algorithm, which is used for short-term load forecasting of integrated energy systems. First, the temporal characteristics of various loads are captured by a temporal multilayer perceptron (MLP), and then the spatial dependencies between variables are modeled by a spatial multilayer perceptron (MLP). Experimental results based on actual ASU-IES operating data show that the proposed method achieves the current advanced level in both prediction accuracy and operating efficiency. Ablation studies show the necessity of temporal and spatial modeling modules and the importance of their combination order. In addition, the significant acceleration effect of the parallel computing strategy is further demonstrated by comparing inference speeds. In summary, the method of this invention achieves a balance between high accuracy and high efficiency while maintaining a low number of parameters, and is suitable for real-time scheduling and operation optimization of modern integrated energy systems. The advantages of this invention are: (1) It constructs a lightweight spatiotemporal architecture based on TSMixer: A lightweight prediction framework is designed for the IES conformity features, in which the temporal pattern within each variable is modeled by an intra-variable multilayer perceptron (MLP), and the spatial dependencies between variables are captured by an inter-variable multilayer perceptron (MLP). While keeping the number of parameters small, it effectively extracts spatiotemporal features. (2) Parallel acceleration strategy for real-time prediction: A parallel computing optimization scheme is proposed to break through the speed bottleneck of the serial structure and significantly reduce the computing time without reducing the accuracy. (3) Balancing high performance and low computational cost: Based on the experimental data of Arizona State University IES, it is verified that the method is superior to advanced models such as CNN, LSTM and Transformer in terms of accuracy, while having higher computational efficiency and smaller model size.

Claims

1. A power grid load fast spatio-temporal prediction method based on a TSMixer algorithm, characterized in that, The method comprises the following steps: S1, processing the original multivariate time series input data from the integrated energy system through an embedding layer to map the original input data to a higher-dimensional latent space; S2, capturing the sequence dependence within each variable by using a time multi-layer perception MLP; S3, transmitting the output of the time multi-layer perception MLP to a spatial multi-layer perception MLP, and the spatial multi-layer perception MLP models the inter-variable correlation between different subsystems; S4, decoding the output of the spatial multi-layer perception MLP to generate the prediction results of all target variables.

2. The method of claim 1, wherein the TSMixer algorithm is based on, The multivariate time series X is mapped to the embedding space by linear projection: (1); where, (2), B represents the data sample size; C represents the number of variables, i.e., the load type; L represents the length of the time series; R represents the data domain; and are learnable parameters.

3. The method of claim 2, wherein the TSMixer algorithm is based on a time-space prediction of power grid load. For , two fully connected layers FC with ReLU activation function are applied: (3); where L denotes the length of the time series; , , W1 （t） and W2 （t） are weight matrices; and are bias terms; is the activation function ReLU; is the hidden dimension of the time feature; the output of the time multi-layer perception MLP is: (4).

4. The method of claim 3, wherein the TSMixer algorithm is based on a time-space prediction of power grid load. The variable vector at a time step is processed as follows: (5); where, , , W1 （S） and W2 （S） are weight s matrices; and are bias terms; is the dimension of the hidden spatial features; and the output of the spatial multi-layer perceptron (MLP) is: (6).

5. The method of claim 4, wherein the TSMixer algorithm is based on a fast spatio-temporal load forecasting method for power grid. mapping the spatial multi-layer perceptron (MLP) processed features to the prediction output by a decoding layer: (7) generating a final multivariate time series prediction by decoding; wherein, W （d） and are learnable parameters; denotes the length of the output sequence.

6. The method of claim 5, wherein the TSMixer algorithm is based on a fast spatio-temporal load forecasting method for power grid. Parallel computation is introduced in time multi-layer perceptron MLP and spatial multi-layer perceptron MLP for optimization: for time multi-layer perceptron MLP, the multivariate time series is adjusted to: (8); all variables are processed in parallel using the same multi-layer perceptron MLP: (9); the output is adjusted to: (10).

Images