Explanatable wind power load space-time prediction method

Abstract

The invention is suitable for the field of wind power load prediction, provides an interpretable wind power load space-time prediction method, considers inherent defects of a recurrent neural network and a traditional attention mechanism and time sequence modes of different spatial data, and improves and designs a novel wind power load prediction method. The problems of gradient, poor parallelism, multiple time sequence scales and the like can be solved in a targeted manner, and a more accurate, efficient and parallel wind power load prediction method is provided for the field of wind power load prediction and a method model library. Aiming at the problem of lack of interpretability of a deep neural network model which is widely researched at present, the wind power load prediction method is combined, and an LIME interpretability method is introduced, so that the wind power load prediction method based on the deep neural network has interpretability.

Description

Technical Field

[0001] The invention belongs to the field of wind power load prediction, and in particular relates to an explainable spatiotemporal prediction method for wind power load. Background Art

[0002] Wind power has gradually become one of the most important clean energy sources due to its environmentally friendly characteristics such as low carbon and no air pollution, and has great development potential. Accurate wind power forecasting has very important practical guiding significance for improving wind power utilization, grid stability, optimizing power dispatching, and optimizing wind power equipment layout. It is also a necessary precursor for wind power grid connection, optimal control, and demand-side management.

[0003] Wind power load forecasting is a typical time series forecasting problem, which establishes the complex nonlinear and spatiotemporal dynamic relationship between wind power load and related parameters through mathematical models based on climate-related parameters, geographic information parameters and historical power generation data of wind farms.

[0004] Common time series forecasting methods can be mainly divided into three types: white box, gray box and black box models. The first two use physical simulation or simplified physical simulation and related historical information to simulate and simulate to make interpretable predictions of wind power load, but they face challenges such as the lack of model capabilities, the complexity of simulation software calculations and the high level of professionalism required by personnel, resulting in large deviations between the predicted values ​​and the actual values. Therefore, black box models represented by machine learning have received widespread attention from researchers. They only need to input historical information and related characteristic variables to accurately and quickly predict historical wind power loads.

[0005] However, both machine learning models and deep learning models can achieve excellent prediction indicators, but they still cannot avoid the unexplainability of typical black box models. At the same time, recurrent neural networks are prone to gradient problems due to the defects of their own structure and have high time complexity. Summary of the invention

[0006] The purpose of the present invention is to provide an interpretable spatiotemporal forecasting method for wind power load, aiming to solve the technical problems existing in the prior art identified in the background technology.

[0007] The present invention is implemented as follows: an interpretable spatiotemporal forecasting method for wind power load, the method comprising:

[0008] Collect wind power data, extract features, preprocess the feature data, and use the random forest algorithm to screen key features of the feature data after data preprocessing;

[0009] Using the temporal convolutional network as the time series model, the TCN is controlled to capture temporal information while extracting the nonlinear features of the data;

[0010] A gated temporal convolutional network is established based on TCN to extract multimodal time series periodic data with different spatial distributions;

[0011] The temporal pattern attention mechanism is introduced to optimize the prediction performance of the time series model by focusing on the information of each time dimension and spatial dimension in the feature space through multiple convolution kernels;

[0012] The dimensioned root mean square error and the dimensionless mean absolute percentage error are defined as the performance evaluation indicators of the model, and the performance of the model is evaluated based on the indicator analysis results.

[0013] As a further solution of the present invention, the use of the random forest algorithm to screen key features of the feature data after data preprocessing specifically includes:

[0014] Assume that the input data is set to X = [x 1 , x 2 , …, x n , y], all features X'=[x 1 , x 2 , …, x n ] is used as the input of the random forest model, and y is used as the target value of the model. The frequency of use of features in each tree on the split node and its Gini non-existence degree on the split node are calculated, and the top N important features required are sorted in descending order to obtain the screened input feature variables. The specific process is as follows:

[0015]

[0016] As a further solution of the present invention, the temporal convolutional network is used as a time series model to control TCN to capture time information while extracting nonlinear features of data, specifically including:

[0017] Collect and organize time series data, and perform missing value processing and normalization;

[0018] Design the input layer of the model and define the shape of the input data, add multiple one-dimensional convolutional layers, and use convolution kernels to extract features from the input data;

[0019] By adjusting the padding mode of the convolution layer to valid, causal convolution is implemented in the convolution layer;

[0020] Set the dilation rate in the convolution layer to expand the receptive field of the convolution operation in the time dimension. Specifically:

[0021] This is achieved by increasing the convolution kernel size k and increasing the dilation rate parameter d. The receptive field size of the nth layer can be calculated using the following formula.

[0022]

[0023] Where k n is the size of the convolution kernel of the nth layer, The accumulated offset for each convolution kernel to expand outward.

[0024] The definition of the convolution operation on sequence elements is as follows:

[0025]

[0026] In the formula, d is the expansion parameter, k is the convolution kernel size, and sd·i represents the direction of convolution;

[0027] The gating strategy is used to weight the results of each head TCN layer to personalize the temporal information extracted by the TCN model at different scales and enhance the diversity of feature extraction. Specifically:

[0028]

[0029]

[0030] X′ t =(A t ) T X t ;

[0031] Where m, l, k m , X′ t , W m , b m , σ, and Γ m They represent the number of GTCN layers, the number of GTCN layers, the output of the m-th TCN in the l-th GTCN layer, the kernel size of the m-th TCN, the model input after feature screening, the corresponding weight parameters, bias, sigmoid activation function, and the mapping between the input and output of the m-th TCN layer;

[0032] Nonlinear transformation is performed after each convolutional layer through activation function;

[0033] Add a residual connection after each convolutional layer to add the input directly to the output after convolution. Specifically:

[0034] o = activation(x+F(x));

[0035] At the last layer of TCN, a fully connected layer is added to map the extracted features to the target output.

[0036] As a further solution of the present invention, the gated temporal convolutional network is established based on TCN to extract multimodal time series periodic data with different spatial distributions, specifically including:

[0037] Construct multiple TCN modules, each module uses a different convolution kernel size and extracts features from different time scales;

[0038] A gating mechanism is introduced at the output of each TCN module to calculate the weight of each output and dynamically adjust the importance of different TCN extracted features in the final output;

[0039] Extract multimodal time series periodic data with different spatial distributions;

[0040] Multiple TCN modules are stacked to form a deep structure to capture more complex temporal dependencies, and the output of each module will serve as the input of the next module;

[0041] After all TCN modules, a fully connected layer is added to map the final features to the target output.

[0042] As a further solution of the present invention, the temporal pattern attention mechanism is introduced to optimize the prediction performance of the time series model by focusing on each time dimension and space dimension information in the feature space through multiple convolution kernels, which specifically includes:

[0043] Use k convolution kernels C i ∈R 1×T Capture the different time patterns of the hidden state Z of the time series model to obtain the output Z corresponding to each hidden layer state and kernel C , specifically:

[0044]

[0045] Then use the scoring function to get Z C The row vectors of are evaluated for correlation and weighted accumulation, specifically:

[0046]

[0047] V t With the last hidden layer state Z t Adaptive weights generate the hidden layer state Z' after attention assistance t , specifically:

[0048] Z′ t =W z Z t +W v V t .

[0049] As a further solution of the present invention, the dimensioned root mean square error and the dimensionless mean absolute percentage error are defined as performance evaluation indicators of the model, and the performance of the model is evaluated based on the indicator analysis results, specifically including:

[0050] Determine the metrics used to evaluate the model's predictive performance, including the dimensioned root mean square error and the dimensionless mean absolute percentage error;

[0051] Calculate the standard deviation of the difference between the true and predicted values, and calculate the coefficient of determination, which measures the goodness of fit between the true and predicted value curves.

[0052] The root mean square error calculation formula is:

[0053]

[0054] The formula for calculating the coefficient of determination is:

[0055]

[0056] in, Represents the model prediction value, y i,j represents the target value, is the number of instances used to calculate the error, l is the output sequence length;

[0057] Make a local linear approximation to the target model, explaining the predictions in the model for specific input data.

[0058] The beneficial effects of the present invention are:

[0059] The present invention takes into account the inherent defects of recurrent neural networks and traditional attention mechanisms and the temporal patterns of different spatial data, improves and designs a new wind power load forecasting method, which can specifically solve problems such as gradient, poor parallelism and multiple time series scales, and provides a more accurate and efficient parallel wind power load forecasting method for the field of wind power load forecasting and method model library.

[0060] In response to the lack of interpretability of the currently widely studied deep neural network models, the present invention combines the proposed wind power load forecasting method with the introduction of the LIME interpretability method to make the proposed wind power load forecasting method based on deep neural network interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 The overall technical scheme diagram of the explainable spatiotemporal forecasting method for wind power load is shown in FIG.

[0062] Figure 2 Schematic diagram of temporal convolutional neural network;

[0063] Figure 3This is a schematic diagram of the gated temporal convolutional network principle;

[0064] Figure 4 Schematic diagram of the principle of the temporal pattern attention mechanism;

[0065] Figure 5 Schematic diagram of the LIME interpretability method. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0067] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0068] Figure 1 The overall technical scheme diagram of the explainable spatiotemporal forecasting method of wind power load is as follows:

[0069] Collect wind power data, extract features, preprocess the feature data, and use the random forest algorithm to screen key features of the feature data after data preprocessing;

[0070] The first step is to use the random forest algorithm to screen the key features of the feature data after data preprocessing. Assume that the input data is set to X = [x1, x2, ..., xn, y], all features X' = [x1, x2, ..., xn] are used as the input of the random forest model, y is used as the target value of the model, and the frequency of use of the features in each tree on the split node and its Gini non-existence degree on the split node are calculated, and the top N important features are screened in descending order to obtain the screened input feature variables. The process is as follows:

[0071]

[0072] In view of the gradient and complexity problems of the recurrent neural network widely used in existing research due to its own recurrent structure, the temporal convolutional neural network (TCN) is introduced as a time series model. It introduces causal convolution, dilated convolution and residual connection on the basis of one-dimensional convolutional neural network (CNN), so that TCN can capture time information while extracting nonlinear features of data. Secondly, considering that wind power loads with different spatial distributions may have temporal periodicity of different scales, we designed a gated temporal convolutional neural network (GTCN) based on TCN to extract multimodal time series periodic data with different spatial distributions. Finally, given that the current research attention mechanism focuses more on the last hidden state of the time dimension in the feature variables extracted by the model, and pays less attention to the spatial dimension and the previous hidden state, the temporal mode attention mechanism is introduced to pay attention to the information of each time dimension and spatial dimension in the feature space through multiple convolution kernels to improve the prediction performance of the model. The relevant composition principles of each module are described as follows:

[0073] Using the temporal convolutional network as the time series model, the TCN is controlled to capture temporal information while extracting the nonlinear features of the data;

[0074] Gated temporal convolutional neural network (GTCN)

[0075] TCN introduces causal convolution and dilated convolution based on traditional one-dimensional CNN, fully considers the information leakage problem in one-dimensional CNN, and provides high parallelism, which can efficiently process time series with low memory consumption. The TCN structure consists of multiple residual blocks, each of which contains two dilated causal convolution layers, which expands the receptive field and enables the model to effectively process long time series. The detailed principle structure is as follows Figure 3 and Figure 4 shown.

[0076] Causal convolution: The causal convolution structure is a hierarchical structure formed by stacking single one-dimensional convolution layers from bottom to top, such as Figure 2 (b) In the causal convolution diagram, the output of the current layer only operates with the elements of the previous layer at an earlier time. Assuming there is a time series data of length T, this sample is used as input and the output value corresponding to each moment is predicted based on this input, the modeling process can be expressed as:

[0077] y 0,y 1 ,...,y T =f(x 0 ,x 1 ,...,x T )

[0078] To make the output y of the model T In relying only on x 0 ,…,x T Under the condition of the value of , we need to find a network f that minimizes the error between the actual value and the predicted output, but the traditional convolutional network obviously cannot do this. Therefore, after some restrictions on the convolution operation, a causal convolution that considers the time sequence is formed. First, the input data enters from the bottom layer, and the time sequence is from left to right. During the convolution process, the output at time t is only convolved with the input at time t, t-1, and t-2, thereby realizing the time sequence convolution.

[0079] Although causal convolution can process time series data well and solve the problem of future information "leaking" into the past caused by traditional convolution operations, when the time series data is too long and the network needs to remember more distant information, a very deep convolution layer is required, which greatly increases the complexity of the model and makes training difficult. In order to solve this problem, the dilated convolution algorithm was proposed and used in conjunction with causal convolution.

[0080] Dilated convolution: The size of the convolution kernel of the dilated convolution remains unchanged, and the number of parameters that need to be updated remains unchanged. The dilated convolution has an additional dilation rate parameter to control the dilation size.

[0081] For example, the receptive field of a 3x3 kernel becomes 7x7 when the expansion rate is 2. Without changing the size of the convolution kernel, increasing the expansion rate parameter and making the receptive field size grow exponentially can make the output results of each layer of convolution contain a wider range of information.

[0082] As shown in the following formula, expanding the receptive field can be achieved by increasing the convolution kernel size k and increasing the dilation rate parameter d. The receptive field size of the nth layer can be calculated using the following formula.

[0083]

[0084] Where k n is the size of the convolution kernel of the nth layer, The accumulated offset for each convolution kernel to expand outward.

[0085] The definition of the convolution operation on sequence elements is as follows:

[0086]

[0087] Where d is the dilation parameter, k is the convolution kernel size, and sd·i represents the direction of convolution.

[0088] Dilated convolution is usually used in conjunction with causal convolution. The schematic diagram of causal convolution is as follows: Figure 2 As shown in (b), causal convolution makes the output at the current moment only focus on the input information of the current and previous moments. At the same time, the dilation rate parameter d in the dilated convolution increases exponentially as the network depth deepens to ensure that very long-term time information can be remembered in the case of a deep network, while alleviating the training difficulty caused by the large amount of computational complexity of the causal convolution operation.

[0089] Residual connection: Increasing the network depth can improve the expressiveness of the model. However, when the network structure is deep, there may be a gradient vanishing phenomenon. The residual connection can solve the problem of gradient vanishing during network training. Temporal convolution (TCN) is stacked layer by layer by residual connection modules. Each residual block contains a branch that is used to learn the transformation function F(x). The output of the transformation function is connected to the input of the next residual block, which can be described as:

[0090] o = activation(x + F(x))

[0091] Figure 2 (a) is the residual connection diagram with kernel size k=3 and dilation factor d=1. The size of the receptive field of TCN depends on the network depth n, the kernel size k and the dilation factor d. Therefore, the network depth and structural complexity of TCN should increase with the increase of long-term memory input information. Each residual block has two dilated causal convolutional layers. The weights are normalized in each dilated causal convolutional layer, and then activated with the regularized linear unit activation function. Finally, a Dropout layer is added for regularization. Since the input and output in TCN may have different numbers of channels, the input cannot be directly added to the output when the number of channels is different. Considering the difference in input and output channels, a one-dimensional convolution structure is added in each residual block to ensure that the residual connection addition layer can receive data with the same shape.

[0092] A gated temporal convolutional network is established based on TCN to extract multimodal time series periodic data with different spatial distributions;

[0093] In order to enhance the ability of TCN to extract temporal information of different scales, GTCN is designed to extract multi-scale features and multi-load data. It consists of multiple TCNs, each equipped with a different kernel size to capture features of different time scales. In addition, a gating strategy is used to weight the results of each head TCN layer to personalize the temporal information extracted by the TCN model at different scales and enhance the diversity of feature extraction. The specific principle can be seen in Figure 3 And the following formula.

[0094]

[0095] X′ t =(A t ) T X t

[0096] Where m, l, km,X′ t , Wm, bm, σ, and Γ m They respectively represent the number of GTCN heads, the number of GTCN layers, the output of the m-th TCN of the l-th layer of GTCN, the kernel size of the m-th TCN, the model input after feature screening, the corresponding weight parameters, bias, sigmoid activation function, and the mapping of the relationship between the input and output of the m-th TCN layer.

[0097] The temporal pattern attention mechanism is introduced to optimize the prediction performance of the time series model by focusing on the information of each time dimension and spatial dimension in the feature space through multiple convolution kernels;

[0098] Temporal pattern attention mechanism (TPA)

[0099] The attention mechanism is to allocate computing resources to more important tasks when computing power is limited, while solving the problem of information overload. The attention mechanism in traditional time series prediction is to focus on the hidden layer state obtained in the time series model [z 1 , z 2 , …, z t ] to obtain more valuable information for prediction performance. However, it only focuses on the last hidden layer state, which makes it impossible to capture the temporal pattern across multiple time steps, making the model prediction performance poor or even weakened. Therefore, this paper adopts the temporal pattern attention mechanism. The principle is as follows Figure 4 As shown in Figure 1, it uses the excellent feature extraction capability of the CNN neural network and uses k convolution kernels Ci∈R1×T to capture the different time patterns of the hidden state Z of the time series model, and obtains the output Z corresponding to each hidden layer state and kernel. C , as shown below:

[0100]

[0101] Then use the scoring function to get Z C The row vectors of are evaluated for correlation and weighted accumulation, as shown in the following formula:

[0102]

[0103]

[0104] Finally, V t With the last hidden layer state Z t Adaptive weights generate the hidden layer state Z' after attention assistance t , such as

[0105] Z′ t =W z Z t +W v V t

[0106] The dimensioned root mean square error and the dimensionless mean absolute percentage error are defined as the performance evaluation indicators of the model, and the performance of the model is evaluated based on the indicator analysis results.

[0107] Model performance evaluation and interpretability analysis: Wind power load forecasting is a typical time series forecasting problem, that is, a regression problem. Therefore, it is necessary to evaluate the quality of the forecasting model based on the difference between the predicted value and the actual value. Based on this, the dimensioned root mean squared error (RMSE) and the dimensionless mean absolute percentage error (MAPE) are determined as the performance evaluation indicators of the model, as described below:

[0108] Error indicator: Root mean square error (RMSE) measures the standard deviation of the difference between the true value and the predicted value. The smaller the value, the smaller the standard deviation of the error between the predicted value and the actual value, and the higher the accuracy of the model. The calculation formula of root mean square error RMSE is shown below.

[0109]

[0110] Coefficient of determination (R 2 ) measures the degree of fit between the true value curve and the predicted value curve. The larger the determination coefficient, the higher the degree of fit between the model prediction value and the target value. 2 ) can be expressed by the following formula.

[0111]

[0112] in, Represents the model prediction value, y i,j represents the target value, is the number of instances used to calculate the error, and l is the output sequence length.

[0113] Interpretability: Model interpretability is an important indicator that cannot be ignored for deep models. Therefore, Local Interpretable Model-agnostic Explanations (LIME) was introduced, which can generate local explanations for any complex model, allowing users to understand the decision-making process of the model, thereby increasing trust in the model output.

[0114] The schematic diagram of LIME related principles is as follows Figure 5 As shown in Figure 1, the core idea is to perform a local linear approximation on the target model (black box model) to explain the prediction of specific input data in the model. Its working principle can be divided into the following steps:

[0115] Generate perturbation data. Based on the given input feature data, generate N perturbation samples to evaluate the local behavior of the model.

[0116] Evaluate the predictions of the perturbed samples, input the perturbed samples into the original black box model, obtain the prediction results for each sample, and use the obtained results to analyze the behavior of the model in this local area.

[0117] Fitting an interpretable model, based on the prediction results of the original black box model, LIME fits a simple model (ridge regression, decision tree, etc.) to these perturbation samples to approximately explain the decision-making process of the original model in the local area.

[0118] Generate local explanations. According to the final LIME fitted model, generate local explanations about the original input samples, as shown in the following formula, which can directly obtain and understand the degree of influence of each feature variable on the model prediction.

[0119]

[0120] In the formula, f, g, G, π x and Ω(g) represent the original model, simple model, set of simple models, and perturbed data set, respectively. The similarity distance with the original dataset X' and the complexity of the simple model.

[0121] In view of the high gradient and time complexity caused by the structural defects of the recurrent neural network itself, as well as the different time periodicity of time series data with different spatial distributions, the present invention proposes an improved gated time convolutional network based on TCN for the extraction of time series information and features of different scales, while increasing the parallelism of the model. Secondly, the attention mechanism used in existing studies acts on the last layer of the hidden state extracted from the time series model, ignoring the role of the previous time step hidden layer in improving the performance of the model. Therefore, on the basis of GTCN, a wind power prediction model based on GTCN and TPA is proposed in combination with the time pattern attention mechanism to comprehensively consider the above-mentioned problems, so as to provide an accurate and parallel wind power load spatiotemporal prediction method, and provide accurate decision support for the subsequent improvement of wind power utilization, grid stability, optimization of power dispatching, and optimization of wind power equipment layout.

[0122] Complex deep learning models are usually regarded as black box models. It is difficult to understand how they make a decision or prediction from the surface, and this lack of transparency leads to deficiencies in the model's credibility. Therefore, the present invention introduces LIME on the basis of the aforementioned wind power spatiotemporal prediction model based on GTCN and TPA to build an interpretable deep wind power load prediction framework, so that the wind power prediction results are interpretable, and wind power management practitioners can further understand the decision-making and prediction process of the model, and increase the trust in the model output.

[0123] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0124] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An interpretable spatiotemporal forecasting method for wind power load, characterized in that: The method comprises: Collect wind power data, extract features, preprocess the feature data, and use the random forest algorithm to screen key features of the feature data after data preprocessing; Using the temporal convolutional network as the time series model, the TCN is controlled to capture temporal information while extracting the nonlinear features of the data; A gated temporal convolutional network is established based on TCN to extract multimodal time series periodic data with different spatial distributions; The temporal pattern attention mechanism is introduced to optimize the prediction performance of the time series model by focusing on the information of each time dimension and spatial dimension in the feature space through multiple convolution kernels; The dimensioned root mean square error and the dimensionless mean absolute percentage error are defined as the performance evaluation indicators of the model, and the performance of the model is evaluated based on the indicator analysis results.

2. The method according to claim 1, characterized in that The use of the random forest algorithm to screen key features of the feature data after data preprocessing specifically includes: Assume that the input data is set to X = [x1, x2, ..., x n , y], all features X'=[x1, x2,…, x n ] is used as the input of the random forest model, and y is used as the target value of the model. The frequency of use of features in each tree on the split node and its Gini non-existence degree on the split node are calculated, and the top N important features required are sorted in descending order to obtain the screened input feature variables. The specific process is as follows:

3. The method according to claim 2, characterized in that The temporal convolutional network is used as a time series model to control TCN to capture time information while extracting nonlinear features of data, specifically including: Collect and organize time series data, and perform missing value processing and normalization; Design the input layer of the model and define the shape of the input data, add multiple one-dimensional convolutional layers, and use convolution kernels to extract features from the input data; By adjusting the padding mode of the convolution layer to valid, causal convolution is implemented in the convolution layer; Set the dilation rate in the convolution layer to expand the receptive field of the convolution operation in the time dimension. Specifically: This is achieved by increasing the convolution kernel size k and increasing the dilation rate parameter d. The receptive field size of the nth layer can be calculated using the following formula. Where k n is the size of the convolution kernel of the nth layer, The accumulated offset for each convolution kernel to expand outward. The definition of the convolution operation on sequence elements is as follows: In the formula, d is the expansion parameter, k is the convolution kernel size, and sd·i represents the direction of convolution; The gating strategy is used to weight the results of each head TCN layer to personalize the temporal information extracted by the TCN model at different scales and enhance the diversity of feature extraction. Specifically: X’ t =(A t ) T X t ; Where m, l, k m , X' t , W m , b m , σ, and Γ m They represent the number of GTCN layers, the number of GTCN layers, the output of the m-th TCN in the l-th GTCN layer, the kernel size of the m-th TCN, the model input after feature screening, the corresponding weight parameters, bias, sigmoid activation function, and the mapping between the input and output of the m-th TCN layer; Nonlinear transformation is performed after each convolutional layer through activation function; Add a residual connection after each convolutional layer to add the input directly to the output after convolution. Specifically: o = activation(x+F(x)); At the last layer of TCN, a fully connected layer is added to map the extracted features to the target output.

4. The method according to claim 3, characterized in that The gated temporal convolutional network is established based on TCN to extract multimodal time series periodic data with different spatial distributions, specifically including: Construct multiple TCN modules, each module uses a different convolution kernel size and extracts features from different time scales; A gating mechanism is introduced at the output of each TCN module to calculate the weight of each output and dynamically adjust the importance of different TCN extracted features in the final output; Extract multimodal time series periodic data with different spatial distributions; Multiple TCN modules are stacked to form a deep structure to capture more complex temporal dependencies, and the output of each module will serve as the input of the next module; After all TCN modules, a fully connected layer is added to map the final features to the target output.

5. The method according to claim 4, characterized in that The temporal pattern attention mechanism is introduced to optimize the prediction performance of the time series model by focusing on the information of each time dimension and space dimension in the feature space through multiple convolution kernels, specifically including: Use k convolution kernels C i ∈R 1×T Capture the different time patterns of the hidden state Z of the time series model to obtain the output Z corresponding to each hidden layer state and kernel C , specifically: Then use the scoring function to get Z C The row vectors of are evaluated for correlation and weighted accumulation, specifically: V t With the last hidden layer state Z t Adaptive weights generate the hidden layer state Z' after attention assistance t , specifically: WITH' t =In z WITH t +W v V t 。 6. The method according to claim 1, characterized in that The dimensioned root mean square error and dimensionless mean absolute percentage error are defined as performance evaluation indicators of the model, and the performance of the model is evaluated based on the indicator analysis results, specifically including: Determine the metrics used to evaluate the model's predictive performance, including the dimensioned root mean square error and the dimensionless mean absolute percentage error; Calculate the standard deviation of the difference between the true and predicted values, and calculate the coefficient of determination, which measures the goodness of fit between the true and predicted value curves. The root mean square error calculation formula is: The formula for calculating the coefficient of determination is: in, Represents the model prediction value, y i,j represents the target value, is the number of instances used to calculate the error, l is the output sequence length; Make a local linear approximation to the target model, explaining the predictions in the model for specific input data.

Images