Energy system multi-load data prediction model construction method and device, storage medium and multi-load data prediction method

Abstract

The invention discloses an energy system multi-load data prediction model construction method and device, a storage medium and a multi-load data prediction method. The model construction method comprises the following steps: step 1, respectively constructing samples to form a data set by using various load and weather characteristic data with relevance meeting requirements generated in each time step in historical time of an energy system, and dividing the data set into a training set, a verification set and a test set; 2, a basic model comprising multiple layers of processing units and an output module is generated, each layer of processing unit comprises a continuous graph learning module, a Mamba network and a bidirectional graph convolution module, the basic model is trained through the training set, and an energy system multi-load data prediction model is obtained. The equipment and the storage medium are used for implementing the model construction method. According to the multi-load data prediction method, various load characteristic data and weather characteristic data of a current time step are collected and input into the energy system multi-load data prediction model, and a load characteristic prediction result is obtained.

Description

Technical Field

[0001] This invention relates to the field of energy system multi-load forecasting methods, specifically a method, device, storage medium, and multi-load data forecasting method for constructing an energy system multi-load data forecasting model. Background Technology

[0002] With traditional fossil fuels such as coal and oil becoming non-renewable and eventually depleted, the cleanliness and reliability of energy supply have become crucial factors affecting national security and sustainable development. The traditional model of separate planning, design, and independent operation of various energy supply and utilization systems (such as electricity, gas, and cooling / heating) significantly reduces energy efficiency. Against this backdrop, the concept of Integrated Energy System (IES) has emerged.

[0003] In integrated energy systems (IES), user energy demand is influenced by various factors such as environmental climate, social development, and human activities. To address this diverse demand, the system integrates different types of energy and energy conversion equipment, such as electric and gas-fired boilers, and adopts innovative methods such as power-to-gas conversion technology and combined cooling, heating, and power (CCHP) systems. Compared to the less than 50% final energy utilization rate of traditional power plants, IES can achieve 70%–90%. This demonstrates that IES achieves mutual conversion and efficient utilization of energy sources, reducing energy waste and pollution emissions, resulting in significant environmental and economic benefits. Load forecasting, as a crucial prerequisite for IES energy management and dispatch optimization, guides the system in the rational dispatch of different energy resources, laying the foundation for the subsequent efficient and stable operation of the system, and thus possesses significant research value.

[0004] Traditional load forecasting refers to electricity load forecasting, that is, predicting the electricity demand for a specific future period. This is the foundation for power system operation, planning, and trading. However, in the context of the energy internet and carbon neutrality, traditional load forecasting can no longer meet the management needs of modern energy systems. Multi-energy load forecasting is a necessary evolution and development direction for traditional load forecasting. It not only needs to predict the independent demand of various energy sources in the future, but also to analyze the coupling, conversion, and substitution relationships between different energy loads.

[0005] Researchers both domestically and internationally have achieved some success in load forecasting, with both traditional statistical methods and deep learning-based approaches proving effective. However, load data exhibits periodicity, trends, and short-term randomness influenced by unforeseen factors. Multi-energy load data retains these characteristics while also possessing multidimensionality and heterogeneity. Furthermore, due to the presence of energy conversion devices, various energy loads exhibit strong coupling and complementarity, and are more strongly influenced by weather factors and social behaviors (such as holidays). Classical statistical forecasting is essentially a linear model, while load and weather factors (such as temperature) have a significant non-linear relationship. Such methods struggle to incorporate high-dimensional external features and have limited capacity to handle massive datasets. Machine learning forecasting methods heavily rely on feature engineering, requiring highly skilled personnel to manually construct features for model comprehension. These methods inherently lack the ability to process sequences. While manually creating features allows the model to "see" data from one or two days prior, this approach is static and limited, making it difficult to flexibly capture long-term dependencies with unpredictable periods. Summary of the Invention This invention provides a method, device, storage medium, and multi-load data prediction method for constructing a multi-load data prediction model for an energy system, in order to solve the problem of poor accuracy in existing multi-load data prediction methods for energy systems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The method for constructing a multi-load data prediction model for energy systems is as follows:

[0008] Step 1: Construct samples based on the various feature data generated at each time step in the history of the energy system. The various feature data include various load feature data and various historical feature data. Thus, each sample contains various load feature data and various historical feature data generated at the corresponding time step.

[0009] A dataset is formed from the individual samples, and then the samples in the dataset are divided into a training set, a validation set, and a test set according to the time order of the time steps.

[0010] Step 2: Generate a basic model, which includes a multi-layer processing unit and an output module. Each processing unit includes a continuous graph learning module, a Mamba network, and a bidirectional graph convolution module.

[0011] The basic model is trained using the training set obtained in step 1, and the trained basic model is used as the energy system multi-load data prediction model.

[0012] In the further step 1, multiple feature datasets generated in the historical time of the energy system are obtained. These multiple feature datasets include multiple load feature datasets and multiple weather feature datasets. Each load feature dataset includes the corresponding type of load feature data generated at each time step in the historical time, and each weather feature dataset includes the corresponding type of weather feature data generated at each time step in the historical time.

[0013] Correlation verification is performed on the feature data in any two different load feature datasets, and correlation analysis is performed on the feature data in any load feature dataset and the feature data in any weather feature dataset. Several weather feature datasets whose correlation analysis results meet the requirements are retained, and the remaining weather feature datasets whose correlation analysis results do not meet the requirements are deleted. Thus, various load feature datasets and various weather feature datasets whose correlation meets the requirements are obtained.

[0014] From various load feature datasets and various weather feature datasets that meet the correlation requirements, multiple load feature data and multiple weather feature data generated at the same time step are respectively used as samples. Each sample includes multiple load feature data and multiple weather feature data generated at the corresponding time step, and the samples form a dataset. Then, the samples in the dataset are divided into training set, validation set and test set according to the time order of the time step.

[0015] Furthermore, the various load characteristics are electrical load characteristics, thermal load characteristics, and cooling load characteristics.

[0016] Furthermore, various weather features were identified, including dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure. Through correlation analysis, seven weather feature datasets were selected, comprising dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure.

[0017] Furthermore, the Pearson correlation coefficient is used to measure linear correlation.

[0018] Furthermore, the Spearman rank correlation coefficient is used to measure nonlinear correlation.

[0019] In further step 2, the data processing procedure for training the base model using the training set obtained in step 1 is as follows:

[0020] Each sample in the training set is input into the continuous graph learning module in each processing unit. The continuous graph learning module performs convolution on each feature in the dataset to obtain the local temporal pattern features of each feature. The continuous graph learning module then maps the local temporal pattern features of each feature data to feature nodes through a weight matrix, and the weight matrix is ​​updated through backpropagation gradient calculation. Subsequently, the continuous graph learning module calculates the similarity between any two feature nodes in the data, learns the correlation between nodes through the weight matrix, and the weight matrix is ​​also updated through backpropagation gradient calculation. Then, the connection score is compressed to between 0 and 1 through the Sigmoid function, thereby obtaining an adjacency matrix composed of multiple feature nodes and edges. Finally, the continuous graph learning module performs sparse smoothing on the obtained continuous adjacency matrix, and uses the generated continuous adjacency matrix to quantify the relationship between multiple load features and multiple weather features.

[0021] The training set sample data is input into the Mamba network in the first processing unit. The Mamba network in the first processing unit outputs a sequence with long-term pattern features to the bidirectional graph convolution module. The bidirectional graph convolution module consists of two graph convolution networks. The first graph convolution network performs graph convolution on the long-term pattern feature sequence using a Laplacian matrix to obtain a forward spatiotemporal feature sequence that integrates long-term pattern features and spatial features. The second graph convolution network performs graph convolution on the long-term pattern feature sequence using a backward Laplacian matrix to obtain a backward spatiotemporal feature sequence. Then, the forward and backward spatiotemporal feature sequences obtained from the two graph convolution networks are accumulated and normalized to obtain the complete spatiotemporal feature sequence. The complete spatiotemporal feature sequence obtained from the first processing unit is finally output to the output module and the Mamba network in the next processing unit.

[0022] In each processing unit after the first layer, the Mamba network performs long-term sequence modeling on the complete spatiotemporal feature sequence obtained by the bidirectional graph convolution module in the previous layer, resulting in a spatiotemporal feature sequence with more complete long-term pattern feature capture. This sequence is then enhanced with spatial features through the bidirectional graph convolution model, resulting in an enhanced spatiotemporal feature sequence. The enhanced spatiotemporal feature sequence is finally output to the output module and the Mamba network in the next processing unit.

[0023] Finally, the spatiotemporal feature sequences of different intensities obtained by the bidirectional graph convolution module in each processing unit are output to the output module. The output module performs residual connection on the spatiotemporal feature sequences obtained by each processing unit, and then performs ReLU activation, one-dimensional convolution, ReLU activation, and one-dimensional convolution in sequence to obtain the prediction results of various load feature data.

[0024] Furthermore, the mean absolute percentage error is used as the loss function during training.

[0025] An electronic device includes a processor and a memory, wherein program instructions in the memory are read and executed by the processor to perform the above-described method for constructing a multi-load data prediction model for an energy system.

[0026] A storage medium storing program instructions, which, when read and executed, perform the above-described method for constructing a multi-load data prediction model for an energy system.

[0027] A method for predicting multiple loads in an energy system, the process of which is as follows:

[0028] Acquire various load data and various weather data generated at the current time step of the energy system;

[0029] Following the above-described method for constructing a multi-load data prediction model for an energy system, a multi-load data prediction model for an energy system is constructed. Multiple load data and multiple weather data generated at the current time step are input into the multi-load data prediction model for an energy system, and the prediction results of various load characteristic data are obtained from the multi-load data prediction model for an energy system.

[0030] Before making predictions, this invention employs Pearson correlation coefficient analysis and Spearman rank correlation coefficient analysis to verify the complex coupling relationships between various feature data. Subsequently, a Mamba-GCN (Graph Convolutional Network) architecture is designed as the base network and trained using the training set obtained after correlation analysis. This graph neural network captures the correlations between different energy sources. Unlike the implicit learning of local patterns in the rule grid structure of convolutional neural networks, it explicitly and efficiently utilizes a known graph structure describing the relationships between features or entities to handle correlations. This invention uses the Mamba network to obtain global information from long-term series to model cross-time-period correlations at a global scale. By combining graph neural networks and the Mamba network, this invention maintains the ability to model long-term series while capturing the coupling relationships between different features, effectively improving the accuracy of multi-energy load prediction.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] 1. This invention proposes a complete correlation analysis method, which verifies the complex coupling relationship between power load, heat load and cooling load from both linear and nonlinear perspectives, and analyzes the correlation between the three and weather factors, thus laying the groundwork for constructing a high-quality feature set.

[0033] 2. This invention proposes and designs a multi-energy load prediction method. By combining Mamba and Graph Convolutional Network (GCN), it can simultaneously learn the relationships within and between different sequences of multi-energy data, effectively enhancing the feature extraction capability of the network and improving the accuracy of prediction. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the method in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the overall architecture of the basic model of the Mamba-GCN architecture in this embodiment of the invention.

[0036] Figure 3 This is a diagram of the Mamba network architecture in an embodiment of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this embodiment discloses a method for constructing a multi-load data prediction model for an energy system. First, linear and nonlinear correlation analyses are performed on three load characteristics (electricity, heat, and cooling) and weather characteristic data using Pearson correlation coefficient analysis and Spearman rank correlation coefficient analysis, respectively, to verify the correlation between these three load characteristics. Simultaneously, weather characteristics with a significant impact on the three load characteristics are identified, thus obtaining training, validation, and test sets. The constructed training set is input into the proposed base model, which includes the Mamba-GCN architecture. The purpose of using the Mamba network in the base model is to capture long-sequence features, as its selective mechanism can effectively model long-term series without incurring the high computational complexity of the Transformer. The purpose of using the graph convolutional module (GCN) in the base model is to explicitly model the relationships between various features, and the graph convolution supplements and corrects the output of the Mamba module, effectively solving the problem of insufficient measurement accuracy caused by inadequate modeling of features.

[0039] The process of constructing the multi-load data prediction model for the energy system in this embodiment is as follows:

[0040] Step 1: In this embodiment, we first acquire multiple feature datasets generated during the historical time of the energy system. These datasets include multiple load feature datasets and multiple weather feature datasets. Each load feature dataset includes corresponding load feature data generated at each time step in the historical time, and each weather feature dataset includes corresponding weather feature data generated at each time step in the historical time.

[0041] Taking a historical period of one year and a time step of one hour as an example, this embodiment acquires three types of load characteristic data and nine types of weather characteristic data generated every day and every hour throughout the year. The three load characteristics are electricity, heat, and cooling. The nine weather characteristics are dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, air pressure, wind direction, and wind speed. This yields three load characteristic datasets and nine weather characteristic datasets.

[0042] In this embodiment, considering the linear relationship between different types of load characteristic data, and the potential nonlinear relationship between multiple loads in the energy system due to the presence of energy conversion devices, correlation verification is performed on the load characteristic data from any two different load characteristic datasets among the three types of load characteristic datasets: electricity, heat, and cooling loads. Specifically, the Pearson correlation coefficient is used to measure the linear correlation between the load characteristic data from any two different load characteristic datasets among the three types of load characteristic datasets. However, the Pearson correlation analysis method only focuses on the linear correlation between variables and cannot capture the nonlinear relationship between variables. Therefore, the Spearman rank correlation coefficient is used to measure the nonlinear correlation between the load characteristic data from any two different load characteristic datasets among the three types of load characteristic datasets. The Spearman rank correlation coefficient is a nonparametric statistical method used to measure the monotonic relationship between two variables. Its core idea is to convert the original data into ranks and then calculate the Pearson correlation coefficient.

[0043] Furthermore, in this embodiment, correlation analysis is performed on weather feature data from any weather feature dataset and load feature data from any of the three load feature datasets. Pearson correlation coefficient is used to measure linear correlation, and Spearman rank correlation coefficient is used to measure nonlinear correlation. Weather feature datasets with both Pearson and Spearman rank correlation coefficients greater than 0.3 with any load feature are selected and retained as those meeting the correlation requirements, while datasets with unsatisfactory results are deleted. In this embodiment, through correlation analysis, seven weather feature datasets—dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure—are ultimately selected and retained from the nine weather feature datasets.

[0044] In this embodiment, the Pearson correlation coefficient is calculated as follows:

[0045]

[0046] In the formula: Covariance is used to measure the degree of common variation between variables x and y, where x and y are any two features. and It indicates the degree of dispersion of variables x and y. and This represents the sample mean of variables x and y. and The selected features are in the first position. The values ​​at each time point.

[0047] In this embodiment, the original data in each feature dataset used to calculate the Spearman rank correlation coefficient is converted into a rank ranking, and then the Pearson correlation coefficient is calculated, which is the Spearman rank correlation coefficient.

[0048] The ranking process involves performing the following steps on the variables: The data is independently sorted in ascending order. Then, a rank is assigned to each data point, numbered 1, 2, 3, and so on, from smallest to largest. Finally, the ranks are placed back into their corresponding positions in the original data to obtain the variables. The ranking of the order.

[0049] The formula for calculating the Spearman rank correlation coefficient is as follows:

[0050]

[0051] In the formula: and Let X and Y represent the rank of variables X and Y, respectively. When calculating the Spearman rank correlation coefficient between any two different load feature datasets, variables X and Y are two different load feature datasets, respectively. When calculating the Spearman rank correlation coefficient between any load feature dataset and any weather feature dataset, one of variables X and Y is any weather feature dataset and the other is a load feature dataset.

[0052] A larger absolute value of the Pearson correlation coefficient indicates a stronger linear correlation between the two variables, and vice versa. Generally speaking, a correlation coefficient of 0.8-1 indicates a very strong correlation, 0.6-0.8 indicates a strong correlation, 0.4-0.6 indicates a moderate correlation, 0.2-0.4 indicates a weak correlation, and 0-0.2 indicates a very weak correlation or no correlation at all. In this embodiment, when calculating the Pearson correlation coefficient and Spearman rank correlation coefficient between each weather feature and any load feature, if both correlation coefficients are greater than 0.3, the corresponding weather feature is retained; otherwise, it is discarded.

[0053] Next, in this embodiment, the three load feature datasets and the seven weather feature datasets that meet the correlation requirements with the load features are used as samples. The three load feature data and the seven weather feature data generated at the same time step are used as samples. Each sample includes the three load feature data (including electricity, heat and cooling) and the seven weather feature data (including dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature and air pressure) generated at the corresponding time step. The dataset is formed from these samples.

[0054] Finally, the samples in the dataset are divided into training set, validation set, and test set according to the time order of the time steps in a ratio of 6:2:2.

[0055] Step 2: Generate a basic model based on the Mamba-GCN architecture. Train the basic model using the training set obtained in Step 1. The trained basic model will then serve as the energy system multi-load data prediction model.

[0056] like Figure 2 As shown, the basic model based on the Mamba-GCN architecture in this embodiment includes a continuous graph learning module, multi-layer processing units, and an output module. Each processing unit includes a Mamba network and a bidirectional graph convolutional network (GCN). During training, the continuous graph learning module generates an adjacency matrix by quantifying the correlation between various feature data for use by the bidirectional graph convolutional network. The Mamba network captures feature relationships over long time sequences. The two graph convolutional networks (GCNs) in the bidirectional graph convolutional network supplement the predictions of the Mamba network through the adjacency matrix. The output of the bidirectional graph convolutional model is output to the output layer and the Mamba network of the next processing unit through residual connections. Finally, the output module outputs the predicted values ​​of electricity, heat, and cooling load features.

[0057] The data processing procedure for training the base model using the training set obtained in step 1 is as follows:

[0058] (1) Each sample in the training set is input into the continuous graph learning module in each layer of the processing unit. The continuous graph learning module generates a continuous adjacency matrix A through dynamic graph structure learning to quantify the relationship between three load features and seven meteorological features, solving the problem that previous models could not accurately capture the complex coupling relationship between multiple load feature data. The correlation values ​​between nodes in the traditional discrete adjacency matrix are only 0 or 1, which can only indicate whether the nodes are related, but cannot indicate the degree of correlation between the nodes. Therefore, this embodiment uses a continuous adjacency matrix to represent the correlation between feature nodes.

[0059] The continuous graph learning module performs convolution on each feature in the dataset. The input time step for each feature is 168, and a convolution kernel of length 12 is used to convolve at each time point of the input step. After all the data at the current time step has been convolved, the data for the next step is input, and so on, finally obtaining the local temporal pattern features of each feature, as shown in the following formula:

[0060]

[0061] In the formula: It is the output of the nth feature at time step j in the i-th time period; k represents the convolution kernel of length k; This indicates that the nth feature is in the i-th time period, and the time step is... The value of .

[0062] The continuous graph learning module then uses a weight matrix to transform the local temporal pattern features of each feature data. Mapped to feature nodes, the weight matrix is ​​updated through backpropagation gradient calculation (i.e., the weight matrix values ​​are updated in the direction of the fastest descent of the network loss function), as shown in the following equation:

[0063]

[0064] In the formula: It is the feature node corresponding to the nth feature. This is the weight matrix. For the local temporal pattern feature of the nth feature, This represents the bias vector, and ReLU is the activation function.

[0065] The continuous graph learning module then calculates the similarity between any two feature nodes in the data. This is done using a weight matrix. The correlation between nodes is learned, and the weight matrix is ​​updated by backpropagation gradient calculation. Finally, the sigmoid function compresses the connection scores to between 0 and 1, thus obtaining the adjacency matrix composed of multiple feature nodes and edges, as shown in the following formula:

[0066]

[0067] In the formula: These are the feature nodes for features i and j, respectively. This is the weight matrix. This represents the bias vector. It is the correlation coefficient between node i and node j in the matrix. This represents the Sigmoid function.

[0068] Finally, the continuous graph learning module performs sparse smoothing on the obtained continuous adjacency matrix. This suppresses weakly correlated connections, keeps moderate connections unchanged, and enhances strong connections, thereby improving model efficiency. The sparse smoothing process is shown in the following equation:

[0069]

[0070] In the formula The sparse smoothing factor is set to 1. Indicates sparse and smooth output; This represents the correlation coefficient between any two nodes.

[0071] (2). The Mamba network is a deep neural network architecture based on the Selective State Space Model (SSM), which is specifically designed for efficient processing of long-term sequences.

[0072] like Figure 3 As shown, in the SSM branch, the Mamba network first performs a linear projection on the input data, expanding the number of channels to obtain multi-channel data (channels not only represent various features, but may also represent a fusion of various features, such as 0.5 temperature - 0.2 humidity representing perceived temperature), allowing the model to learn richer and more expressive temporal features. Subsequently, a 1×1 convolution is performed on different channels at the same time point of the multi-channel data, weighted and summed to obtain the interaction relationships between different channels. The convolution result is then nonlinearly introduced through the SiLU activation function, enhancing the model's ability to fit complex patterns. The data is then input into the core SSM of the Mamba network to capture long sequence dependencies and dynamically select important information. The other branch undergoes linear transformation and lossless transfer using the SiLU function, ensuring model stability and reliability and preventing the loss of key original information in deep networks. Finally, the data from the two branches are fused, compressed back to the original number of channels through linear projection, and output.

[0073] During the discretization of continuous SSM The time step scaling factor controls the frequency of state updates. When the value of B is large, state updates are slower and historical information is retained for longer; when the value of B is small, state updates are more frequent and more recent inputs are considered. Another part of the selective function of the SSM relies on gating components. The system matrix B acts as the information input gating; when B is large, the current input is strongly written into the state, i.e., important information; when B approaches 0, the current input is ignored and judged as invalid information. The output gating is the system matrix C; similarly, a large value strongly outputs information, while a small value suppresses output.

[0074] First, SSM discretizes the data after the linear projection layer normalizes the convolutional layer features and then performs a linear projection on the convolutional input to obtain the discretization stride parameter. Input projection parameters and output projection parameters The splicing method, split, divides the splice into a dissociation step size parameter. Input projection parameters (i.e., input projection matrix B) and output projection parameters (i.e., the output projection matrix C), as shown in the following formula:

[0075]

[0076] In the formula: Discretization step size parameter These are the input projection parameters. Segmentation operation, It is a linear projection. It is the input feature processed by one-dimensional convolution. It outputs the projection parameters.

[0077] Subsequently, the logarithmic form of the state transition matrix A after structure initialization is discretized using the discretization step size parameter, and then restored from the logarithmic form to obtain the discrete form of the state transition matrix A. The A matrix can be updated by back-calculating the parameters in the update matrix using gradients, as shown in the following equation:

[0078]

[0079] In the formula It is a discretized state transition matrix. It is the logarithmic parameter of the continuous-time state matrix.

[0080] Finally, the input projection matrix B is discretized. The product of the matrix parameters and the discretization step size parameter yields the discretized input matrix, as shown in the following equation:

[0081]

[0082] In the formula: It is a discretized input matrix.

[0083] After discretizing the state transition matrix and input projection matrix, the system enters the state recursion stage, iteratively deriving the optimal BC matrix parameters of the model. The SSM in the Mamba block is essentially a recursive system, and its recursive state equations and outputs are shown below:

[0084]

[0085] In the formula: Let be the hidden state at time t. It is the hidden state at time t-1. Let be the discrete state transition matrix at time t. Let be the output projection matrix at time t. The input data is at time t.

[0086] After recursively deriving the optimal parameters, the output is given. The D matrix is ​​a direct mapping matrix from input to output in the state-space model, bypassing the state evolution process. It is used to ensure the stability of the system and the instantaneous response capability of the SSM. The D matrix is ​​updated using the gradient descent algorithm, as shown in the following equation:

[0087]

[0088] In the formula: This is the final output of SSM. To output the projection matrix, It is a directly connected matrix.

[0089] In this embodiment, the sample data of the training set is input into the Mamba network in the first layer processing unit. The Mamba network in the first layer processing unit outputs a sequence with long-term pattern features to the bidirectional graph convolution module. The bidirectional graph convolution module includes two graph convolution networks. The first graph convolution network performs graph convolution processing on the long-term pattern feature sequence using the Laplacian matrix (calculated from the adjacency matrix of the continuous graph learning input) to obtain a positive spatiotemporal feature sequence that integrates long-term pattern features and spatial features (referring to the correlation between features). The symmetric normalized Laplacian calculation and convolution process are as follows:

[0090]

[0091] Where A is the adjacency matrix generated by the continuous graph learning module, and D is the degree matrix calculated from matrix A. It is a symmetric normalized Laplace matrix.

[0092] In graph convolution, the weight matrix W learns how to map the spatially aggregated features to the optimal prediction target through the gradient descent algorithm, as shown in the following equation:

[0093]

[0094] In the formula: Let A be a symmetric normalized Laplacian matrix, I be the adjacency matrix, D be the degree matrix, and W be the weight matrix. It is the ReLU activation function; It is a spatiotemporal feature sequence; It is a long-term pattern feature sequence.

[0095] The second graph convolutional network uses the inverse Laplacian matrix to perform graph convolution on the long-term pattern feature sequence to obtain the inverse spatiotemporal feature sequence.

[0096] Then, the forward and backward spatiotemporal feature sequences obtained from the two graph convolutional networks are accumulated and normalized to obtain the complete spatiotemporal feature sequence. The complete spatiotemporal feature sequence obtained from the first processing unit is finally output to the output module and the Mamba network in the next processing unit.

[0097] Bidirectional graph convolution is implemented by transposing the adjacency matrix A and inputting it into a second graph convolution module. Using two independent graph convolutions to process forward and reverse information respectively, the two directions complement each other, enabling a more accurate learning of complex relationships between features. This also serves as an implicit regularization, helping to prevent overfitting and promoting the model to learn more robust feature representations.

[0098] (4) In each processing unit after the first processing unit, the Mamba network performs long-term sequence modeling on the complete spatiotemporal feature sequence obtained by the bidirectional graph convolution module in the previous processing unit to obtain a spatiotemporal feature sequence with more complete long-term pattern feature capture. Then, the spatial feature is enhanced by the bidirectional graph convolution model to obtain the enhanced spatiotemporal feature sequence. The enhanced spatiotemporal feature sequence is finally output to the output module and the Mamba network in the next processing unit.

[0099] (5) Finally, the spatiotemporal feature sequences of different intensities obtained by the bidirectional graph convolution module in each processing unit are output to the output module; the output module performs residual connection on the spatiotemporal feature sequences obtained by each processing unit, and then performs ReLU activation, one-dimensional convolution, ReLU activation, and one-dimensional convolution in sequence to obtain the prediction results of various load feature data.

[0100] The basic model is trained using a training set to obtain a well-trained basic model, which serves as a prediction model for multi-load data in the energy system.

[0101] In this embodiment, the mean absolute percentage error (MASE) is used as the loss function during training. Before training, the basic model parameters are initialized, and the learning rate and decay rate are determined. Specifically, the learning rate is set to 0.0001, the number of inputs per training iteration is set to 64, the time step is set to 168, and a total of 100 training iterations are performed. The Adam optimizer is used to train the neural network and update the model parameters.

[0102] The training data is input into the network model for training, hyperparameters are adjusted, and training is performed multiple times. Then, the validation set is input into the trained model for validation. The mean absolute percentage error (MAE) is calculated between the output results and the labeled data. The model with the smallest MAE is the energy system multi-load data prediction model.

[0103] The trained energy system multi-load data prediction model was tested on the test set data. The output results were compared with the actual load values. Three evaluation indicators were output: mean absolute error, mean absolute percentage error, and root mean square error. The performance of the proposed network was compared with that of other existing mainstream models to observe its performance.

[0104] This embodiment also uses the obtained energy system multi-load data prediction model to predict the publicly available multi-energy load data of Arizona State University Tempe Campus, and provides a multi-energy load prediction process based on the Mamba-GCN network, as follows:

[0105] After performing correlation analysis on load and weather data, weather features with a correlation coefficient greater than 0.3 for the three load types were retained. Dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure were selected as auxiliary features from the weather data. Holidays and weekends also affect the load; therefore, holidays (1 for holidays, 0 for non-holidays), weekends (1 for weekends, 0 for non-weekends), month, day, and hour were selected as auxiliary features from the calendar information. The feature set was then input into the Mamba-GCN network for prediction. To verify the effectiveness of the prediction model presented in this paper, four conventional deep learning models were trained: GRU, Transformer, CNN-LSTM, and CNN-GRU-Attention, as shown in Table 1.

[0106] Table 1 Network Performance Comparison Table

[0107]

[0108] The Mamba-GCN in Table 1 is the energy system multi-load data prediction model obtained in this embodiment. As can be seen from Table 1, the MAPE of the energy system multi-load data prediction model obtained in this embodiment is much smaller than that of other comparative models, effectively improving the accuracy of multi-energy load prediction.

[0109] This embodiment also discloses an electronic device, including a processor and a memory, wherein program instructions in the memory are read and executed by the processor to perform steps 1 and 2 of the above-described method for constructing a multi-load data prediction model for an energy system.

[0110] This embodiment also discloses a storage medium that stores program instructions. When the program instructions are read and run, steps 1 and 2 of the above-described method for constructing a multi-load data prediction model for an energy system are executed.

[0111] This embodiment also discloses a method for predicting multiple load data in an energy system, the process of which is as follows:

[0112] Acquire three types of load characteristic data (including electricity, heat and cooling) and seven types of weather characteristic data (including dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature and air pressure) generated by the energy system at the current time step.

[0113] Following steps 1 and 2 of the above-mentioned method for constructing a multi-load data prediction model for an energy system, a multi-load data prediction model for an energy system is constructed. Multiple load data and multiple weather data generated at the current time step are input into the multi-load data prediction model for an energy system, and the prediction results of three load characteristic data—electricity, heat, and cooling—are obtained from the multi-load data prediction model for an energy system.

[0114] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.

[0115] This invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this invention and without departing from the design idea of ​​this invention, all modifications and improvements made by those skilled in the art to the technical solutions of this invention should fall within the protection scope of this invention. The technical content for which protection is sought in this invention has been fully described in the claims.

Claims

1. A method for constructing a multi-load data prediction model for an energy system, characterized in that, The process is as follows: Step 1: Construct samples based on the various feature data generated at each time step in the history of the energy system. The various feature data include various load feature data and various historical feature data. Thus, each sample contains various load feature data and various historical feature data generated at the corresponding time step. A dataset is formed from the individual samples, and then the samples in the dataset are divided into a training set, a validation set, and a test set according to the time order of the time steps. Step 2: Generate a basic model, which includes a multi-layer processing unit and an output module. Each processing unit includes a continuous graph learning module, a Mamba network, and a bidirectional graph convolution module. The basic model is trained using the training set obtained in step 1, and the trained basic model is used as the energy system multi-load data prediction model.

2. The method for constructing a multi-load data prediction model for an energy system according to claim 1, characterized in that, In step 1, multiple feature datasets generated in the historical time of the energy system are obtained. These multiple feature datasets include multiple load feature datasets and multiple weather feature datasets. Each load feature dataset includes the corresponding type of load feature data generated at each time step in the historical time, and each weather feature dataset includes the corresponding type of weather feature data generated at each time step in the historical time. Correlation verification is performed on the feature data in any two different load feature datasets, and correlation analysis is performed on the feature data in any load feature dataset and the feature data in any weather feature dataset. Several weather feature datasets whose correlation analysis results meet the requirements are retained, and the remaining weather feature datasets whose correlation analysis results do not meet the requirements are deleted. Thus, various load feature datasets and various weather feature datasets whose correlation meets the requirements are obtained. From various load feature datasets and various weather feature datasets that meet the correlation requirements, multiple load feature data and multiple weather feature data generated at the same time step are respectively used as samples. Each sample includes multiple load feature data and multiple weather feature data generated at the corresponding time step, and the samples form a dataset. Then, the samples in the dataset are divided into training set, validation set and test set according to the time order of the time step.

3. The method for constructing a multi-load data prediction model for an energy system according to claim 1 or 2, characterized in that, The various load characteristics are electrical load characteristics, thermal load characteristics, and cooling load characteristics.

4. The method for constructing a multi-load data prediction model for an energy system according to claim 2, characterized in that, The dataset includes various weather features such as dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure. Through correlation analysis, seven weather feature datasets were selected, including dew point temperature, precipitation, solar zenith angle, relative humidity, surface albedo, temperature, and air pressure.

5. The method for constructing a multi-load data prediction model for an energy system according to claim 2, characterized in that, The Pearson correlation coefficient is used to measure linear correlation.

6. The method for constructing a multi-load data prediction model for an energy system according to claim 2, characterized in that, Spearman's rank correlation coefficient is used to measure nonlinear correlation.

7. The method for constructing a multi-load data prediction model for an energy system according to claim 1, characterized in that, In step 2, the data processing procedure for training the base model using the training set obtained in step 1 is as follows: Each sample in the training set is input into the continuous graph learning module in each layer of the processing unit. The continuous graph learning module performs convolution on each feature in the dataset to obtain the local temporal pattern features of each feature. The continuous graph learning module then maps the local temporal pattern features of each feature data to feature nodes through a weight matrix, and the weight matrix is ​​updated through backpropagation gradient calculation. Subsequently, the continuous graph learning module calculates the similarity between any two feature nodes in the data, learns the correlation between nodes through the weight matrix, and the weight matrix is ​​also updated through backpropagation gradient calculation. Then, the connection score is compressed to between 0 and 1 through the Sigmoid function, thereby obtaining an adjacency matrix composed of multiple feature nodes and edges. Finally, the continuous graph learning module performs sparse smoothing on the obtained continuous adjacency matrix, and uses the generated continuous adjacency matrix to quantify the relationship between various load features and various weather features. The training set sample data is input into the Mamba network in the first processing unit. The Mamba network in the first processing unit outputs a sequence with long-term pattern features to the bidirectional graph convolution module. The bidirectional graph convolution module consists of two graph convolution networks. The first graph convolution network performs graph convolution on the long-term pattern feature sequence using a Laplacian matrix to obtain a forward spatiotemporal feature sequence that integrates long-term pattern features and spatial features. The second graph convolution network performs graph convolution on the long-term pattern feature sequence using a backward Laplacian matrix to obtain a backward spatiotemporal feature sequence. Then, the forward and backward spatiotemporal feature sequences obtained from the two graph convolution networks are accumulated and normalized to obtain the complete spatiotemporal feature sequence. The complete spatiotemporal feature sequence obtained from the first processing unit is finally output to the output module and the Mamba network in the next processing unit. In each processing unit after the first layer, the Mamba network performs long-term sequence modeling on the complete spatiotemporal feature sequence obtained by the bidirectional graph convolution module in the previous layer, resulting in a spatiotemporal feature sequence with more complete long-term pattern feature capture. This sequence is then enhanced with spatial features through the bidirectional graph convolution model, resulting in an enhanced spatiotemporal feature sequence. The enhanced spatiotemporal feature sequence is finally output to the output module and the Mamba network in the next processing unit. Finally, the spatiotemporal feature sequences of different intensities obtained by the bidirectional graph convolution module in each processing unit are output to the output module. The output module performs residual connection on the spatiotemporal feature sequences obtained by each processing unit, and then performs ReLU activation, one-dimensional convolution, ReLU activation, and one-dimensional convolution in sequence to obtain the prediction results of various load feature data.

8. The method for constructing a multi-load data prediction model for an energy system according to claim 7, characterized in that, The mean absolute percentage error is used as the loss function during training.

9. An electronic device comprising a processor and a memory, characterized in that, The program instructions in the memory are read and executed by the processor to perform the energy system multi-load data prediction model construction method as described in any one of claims 1-8.

10. A storage medium storing program instructions, characterized in that, When the program instructions are read and executed, the method for constructing a multi-load data prediction model for an energy system as described in any one of claims 1-8 is performed.

11. A method for predicting multiple loads in an energy system, characterized in that, The process is as follows: Acquire various load data and various weather data generated at the current time step of the energy system; According to the method for constructing a multi-load data prediction model for an energy system as described in any one of claims 1-8, a multi-load data prediction model for an energy system is constructed; multiple load data and multiple weather data generated at the current time step are input into the multi-load data prediction model for an energy system, and the prediction results of various load characteristic data are obtained from the multi-load data prediction model for an energy system.

Images