Fourier sequence diagram learning method and system for cross-cycle power load prediction

Abstract

The invention discloses a Fourier sequence diagram learning method and system for cross-cycle power load prediction, and relates to the technical field of power system prediction, the system comprises a feature extraction module, a spatial relationship modeling module and a data processing and feature fusion module, and multi-scale Fourier transform belongs to the feature extraction module. Parallel graph attention network expansion belongs to a spatial relation module, joint optimization framework perfection belongs to a data processing and feature fusion module, multi-scale Fourier transform comprises dynamic spectrum reconstruction and multi-resolution spectrum fusion, the dynamic spectrum reconstruction is based on a recurrent neural network and variants thereof, and the multi-resolution spectrum fusion is based on the recurrent neural network. According to the method, the future change trend is predicted according to the fluctuation mode in the historical data, the frequency domain features are dynamically adjusted, and the dynamic change of the data is captured, and the method has the advantages that the power load is predicted more accurately by using the multi-scale Fourier transform and the parallel graph attention network.

Description

Technical Field

[0001] The present invention relates to the technical field of power system prediction, and in particular to a Fourier time series graph learning method and system for cross-cycle power load prediction. Background Art

[0002] In modern power systems, with the rapid development of the economy and society and the widespread use of various new types of electrical equipment, power loads have become complex and changeable. Accurate power load forecasting is crucial for the safe and stable operation of power systems, the optimal allocation of power resources, and the efficient operation of power markets. Power load forecasting must not only consider short-term load fluctuations but also cross-cycle trends to cope with load changes on different time scales, such as seasonal changes and the difference between weekdays and holidays. As a powerful signal analysis tool, Fourier transform can convert time domain signals into frequency domain signals and clearly display the periodic components of the signal. Combining Fourier transform with deep learning and constructing a learning method based on Fourier time series graphs is expected to provide a new solution for cross-cycle power load forecasting. By performing Fourier transform on power load data, its characteristics at different frequencies can be obtained, and then these characteristics can be learned and predicted using deep learning models. To this end, we propose a Fourier time series graph learning method and system for cross-cycle power load forecasting. Summary of the Invention

[0003] The object of the present invention is to provide a Fourier time series graph learning method and system for cross-cycle power load forecasting.

[0004] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a Fourier time series graph learning system for cross-cycle power load forecasting, including a feature extraction module, a spatial relationship modeling module, and a data processing and feature fusion module. The multi-scale Fourier transform belongs to the feature extraction module, the parallel graph attention network extension belongs to the spatial relationship module, and the joint optimization framework improvement belongs to the data processing and feature fusion module. The multi-scale Fourier transform includes dynamic spectrum reconstruction and multi-resolution spectrum fusion. The parallel graph attention network extension includes the introduction of reinforcement learning to dynamically adjust weights and the fusion of multi-source heterogeneous data to enhance spatial dependency expression. The joint optimization framework improvement includes data enhancement of generative adversarial networks and adaptive multimodal feature fusion. The dynamic spectrum reconstruction is based on recursive neural networks and their variants. It predicts future change trends based on fluctuation patterns in historical data, dynamically adjusts frequency domain features and captures dynamic changes in data, and combines the lonely forest algorithm and density-based spatial clustering algorithm to identify and eliminate outliers.

[0005] As a further solution of the present invention: the dynamic spectrum reconstruction in the multi-scale Fourier transform is to establish a dynamic model to monitor the change trend of the data in real time after Fourier transforming the power load data to obtain frequency domain features. The dynamic model is constructed based on a recursive neural network and its variants, and the fluctuation patterns in the historical data are collected and stored in the recursive neural network. The future change trend is predicted based on the historical data. For the power load data point, its outlier score in the lonely forest is calculated. When it is greater than the set threshold, it is determined to be an outlier. An interpolation model based on a bidirectional long short-term memory network is used. Assume that the power load time series has a missing value position of , and the missing value is predicted using a Bi-LSTM model: The forward propagation formula of the Bi-LSTM model is: Forget Gate: ; Input Gate: ; Candidate memory cells: ; Memory unit: ; Output gate: ; Hidden state: ; Bidirectional hidden state forward and backward Predict missing values:

[0006] in, as well as is the weight matrix, as well as is the bias vector, is the Sigmoid function, is element-wise multiplication, represents the data input to the Bi-LSTM model at time step t, is the hidden state of the Bi-LSTM model at time step t-1, is the memory cell at time step t-1, and j represents the position where the value is missing in the power load time series; Data normalization uses the Min-Max normalization method, and the specific formula is:

[0007] Normalize the power load data to interval, Represents the raw power load data.

[0008] As a further solution of the present invention, the multi-resolution spectrum fusion in the multi-scale Fourier transform specifically performs Fourier transform on the daily, weekly and monthly data respectively to obtain the corresponding frequency domain features, and designs a hierarchical fusion algorithm to cross-fuse the obtained frequency domain features, and divide the original power load data into multi-scale time series, and the daily scale series, weekly scale series and monthly scale series are respectively:

[0009] in, is the total number of time steps, one day is time steps, one week time steps, one month time steps; According to the dynamic spectrum reconstruction, the time series of each scale are transformed by Fourier transform to obtain the frequency domain features. Based on the gated cyclic unit, a dynamic model is constructed to detect the data change trend. The input of the gated cyclic unit is set as the frequency domain feature sequence. , the hidden state is ,but: Update Gate: Reset the gate: Candidate hidden states: Hidden state: Survey coefficients for predicting future frequency domain feature changes based on hidden states , the adjusted frequency domain features ,in Depend on Through the fully connected layer, we can get: ; in is an element in the frequency domain feature sequence at time step t, represents the hidden state at time step t-1, Based on the hidden state Survey coefficients obtained through the fully connected layer.

[0010] As a further solution of the present invention: the introduction of reinforcement learning to dynamically adjust weights in the expansion of the parallel graph attention network is to introduce a reinforcement learning mechanism on the basis of the parallel graph attention network to dynamically adjust the node relationship weights, regard the power load prediction task as a reinforcement learning problem, and define the agent, environment, state, action and reward. The agent is the weight adjustment module in the parallel graph attention network, the environment is the power load data and its related spatial relationship information, the state is the power load data characteristics and graph structure at the current moment, the action is the adjustment operation of the node relationship weights, and the reward is determined based on the error between the predicted structure and the actual power load.

[0011] As a further solution of the present invention: the fusion of multi-source heterogeneous data in the expansion of the parallel graph attention network enhances the expression of spatial dependency by fusing meteorological data and geographic information data. The meteorological data and geographic information data are first subjected to feature engineering processing. The meteorological data includes temperature, humidity and wind speed, and the geographic information data includes regional functional division, population density and topography. For meteorological data, the sliding average and wavelet transform methods are used to extract its time series features. For geographic information data, encoding and normalization operations are used to convert it into a numerical feature vector, and the processed multi-source heterogeneous data is fused with the power load data. In the parallel graph attention network, the fused features are used as the input features of the nodes to participate in the generation process of the relationship matrix.

[0012] As a further solution of the present invention: the data enhancement of the generative adversarial network in the improvement of the joint optimization framework is to enhance the power load data by using the generative adversarial network in the data preprocessing stage. The generative adversarial network is composed of a generator and a discriminator. The generator is used to learn the distribution characteristics of the original power load data and generate synthetic data that is the same as the real data. The generator uses a deep convolutional neural network and a variational autoencoder to input a noise vector and output synthetic power load data. The discriminator is responsible for distinguishing between real data and synthetic data, and it uses a multi-layer perceptron structure.

[0013] As a further solution of the present invention: The adaptive multimodal feature fusion in the improvement of the joint optimization framework adopts an adaptive multimodal feature fusion method in the process of spatiotemporal feature fusion. Based on the attention mechanism, the fusion weights of the modal features are automatically adjusted according to the data characteristics of the time scale and spatial position, the modal features are respectively input into their respective feature extraction networks, the feature vectors are extracted, and the correlation between each feature vector and the target prediction value is calculated.

[0014] As a further solution of the present invention: the Bi-LSTM model uses a bidirectional structure and a gating mechanism to learn the long-term dependencies of the data based on the data information of the previous and next moments in the power load time series, and effectively predicts missing values. In the process of predicting missing values, the Bi-LSTM model deeply mines the time series characteristics of the power load data and captures the trends and patterns in the data. The feature extraction module provides processed multi-scale frequency domain features to the spatial relationship modeling module and the data processing and feature fusion module. The spatial relationship modeling module uses features combined with reinforcement learning and multi-source heterogeneous data to construct a spatial relationship model.

[0015] A Fourier time series graph learning method for cross-cycle power load forecasting includes the following steps: data processing, feature extraction, spatial relationship modeling, data enhancement, and feature fusion. The isolation forest algorithm and DBSCAN are combined to identify and remove outliers in the power load data. For missing values, the Bi-LSTM model is used to predict and fill in the missing values ​​based on its bidirectional structure and gating mechanism to learn the long-term dependency of the data. The Min-Max normalization method is used to normalize the data to Interval, divide the original power load data into daily, weekly and monthly granularity to construct multi-scale time series, use multi-scale Fourier transform, dynamic spectrum reconstruction and multi-resolution spectrum fusion to extract features; The parallel graph attention network is expanded, reinforcement learning is introduced to dynamically adjust the node relationship weights, the prediction task is regarded as a related element of the reinforcement learning problem definition, and multi-source heterogeneous data such as meteorological and geographic information are integrated. After feature engineering processing, they participate in the generation of relationship matrices, build spatial relationship models, and improve the joint optimization framework. The generative adversarial network is used to enhance data in the data preprocessing stage. The generator generates synthetic data, and the discriminator distinguishes between true and false. An adaptive multimodal feature fusion method is adopted. Based on the attention mechanism, the fusion weights of different modal features are adjusted according to the characteristics of time and space data to improve prediction accuracy.

[0016] By adopting the above technical solution, compared with the prior art, the beneficial effects of the present invention are: 1. By using a bidirectional long short-term memory network model, the present invention can fully utilize the contextual information of the power load time series to more accurately predict and fill missing values. It can capture the long-term dependencies and dynamic changes in power load data, thereby more accurately estimating missing values. This provides a high-quality data foundation for subsequent operations such as multi-scale Fourier transform, parallel graph attention network expansion, and joint optimization framework improvement, thereby improving the accuracy and reliability of the entire power load forecasting system. 2. Through dynamic spectrum reconstruction and multi-resolution spectrum fusion technology, the present invention can deeply mine the short-term fluctuations, medium-term trends and long-term cycle characteristics in power load data. At the same time, it can dynamically adjust the frequency domain characteristics to adapt to the real-time changes in power load, comprehensively and accurately grasp the multi-scale characteristics, and provide a solid data foundation for accurate prediction. By expanding the parallel graph attention network and introducing reinforcement learning to dynamically adjust weights, it can dynamically optimize the node relationship weights based on real-time power load data and spatial relationship information, and accurately characterize the spatial dependency pattern between users. At the same time, the integration of multi-source heterogeneous data such as meteorological data and geographic information data further enhances the expression ability of spatial dependency relationships, enabling the model to more accurately reflect the differences in electricity consumption in different regions and under different environmental conditions, effectively handle spatiotemporal dependencies, and improve the accuracy of predictions. 3. The present invention identifies and eliminates outliers by combining the isolation forest algorithm and the density-based spatial clustering algorithm, predicts and fills missing values ​​using an interpolation model based on a bidirectional long short-term memory network, and standardizes the data using the Min-Max normalization method. In addition, the generative adversarial network data enhancement technology in the joint optimization framework can expand the diversity of data, so that the model has stronger adaptability and generalization ability when facing complex and changeable data. Even when data is scarce or noisy, it can still maintain a high prediction accuracy. Through the adaptive multimodal feature fusion method, based on the attention mechanism, the fusion weights of the modal features are automatically adjusted according to the data characteristics of the time scale and spatial position, which can effectively focus on the features that contribute more to the prediction results, suppress the influence of interference features, and achieve efficient fusion of multimodal features. This enables the model to more accurately learn the complex patterns in the power load data, significantly improve the prediction performance, and provide a more reliable decision-making basis for the planning, scheduling and management of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flowchart of the Fourier timing diagram learning system in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0019] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Please see the attached Figure 1The Fourier time series graph learning system for cross-cycle power load prediction of the present invention includes a feature extraction module, a spatial relationship modeling module, and a data processing and feature fusion module. The multi-scale Fourier transform belongs to the feature extraction module, the parallel graph attention network expansion belongs to the spatial relationship module, and the joint optimization framework improvement belongs to the data processing and feature fusion module. The multi-scale Fourier transform includes dynamic spectrum reconstruction and multi-resolution spectrum fusion. The parallel graph attention network expansion includes the introduction of reinforcement learning to dynamically adjust weights and the fusion of multi-source heterogeneous data to enhance spatial dependency expression. The joint optimization framework improvement includes data enhancement of the generative adversarial network and adaptive multimodal feature fusion. The dynamic spectrum reconstruction is based on recursive neural networks and their variants. According to the fluctuation pattern in historical data, the future change trend is predicted, the frequency domain features are dynamically adjusted and the dynamic changes of the data are captured. At the same time, the lonely forest algorithm and the density-based spatial clustering algorithm are combined to identify and eliminate outliers.

[0021] In one embodiment of the present invention, dynamic spectrum reconstruction in multi-scale Fourier transform is to establish a dynamic model to monitor the change trend of data in real time after Fourier transforming the power load data to obtain frequency domain features. The dynamic model is constructed based on recursive neural networks and their variants, and the fluctuation patterns in historical data are collected and stored in the recursive neural network. The future change trend is predicted based on the historical data. For power load data points, their outlier scores in the lonely forest are calculated. When the score is greater than the set threshold, it is determined to be an outlier. An interpolation model based on a bidirectional long short-term memory network is used. Assume that the power load time series has a missing value position and the Bi-LSTM model is used to predict the missing value: The forward propagation formula of the Bi-LSTM model is: Forget Gate: Input Gate: Candidate memory cells: Memory unit: Output gate: Hidden state: Bidirectional hidden state forward and backward Predict missing values:

[0022] in, as well as is the weight matrix, as well as is the bias vector, is the Sigmoid function, is element-wise multiplication, represents the data input to the Bi-LSTM model at time step t, is the hidden state of the Bi-LSTM model at time step t-1, is the memory cell at time step t-1, and j represents the position where the value is missing in the power load time series; Data normalization uses the Min-Max normalization method, and the specific formula is:

[0023] Normalize the power load data to interval, Represents the raw power load data.

[0024] In one embodiment of the present invention, the multi-resolution spectrum fusion in the multi-scale Fourier transform specifically performs Fourier transform on the daily, weekly and monthly data respectively to obtain the corresponding frequency domain features, and designs a hierarchical fusion algorithm to cross-fuse the obtained frequency domain features. The original power load data is divided into multi-scale time series, and the daily scale series, weekly scale series and monthly scale series are respectively:

[0025] in, is the total number of time steps, one day is Time step, one week time steps, one month time steps; According to the dynamic spectrum reconstruction, the time series of each scale are transformed by Fourier transform to obtain the frequency domain features. Based on the gated cyclic unit, a dynamic model is constructed to detect the data change trend. The input of the gated cyclic unit is set as the frequency domain feature sequence. , the hidden state is ,but: Update Gate: Reset the gate: Candidate hidden states: Hidden state: Survey coefficients for predicting future frequency domain feature changes based on hidden states , the adjusted frequency domain features ,in Depend on Through the fully connected layer, we can get: ; in is an element in the frequency domain feature sequence at time step t, represents the hidden state at time step t-1, Based on the hidden state Survey coefficients obtained through the fully connected layer.

[0026] In one embodiment of the present invention: the introduction of reinforcement learning in the expansion of the parallel graph attention network to dynamically adjust the weights is to introduce a reinforcement learning mechanism on the basis of the parallel graph attention network to dynamically adjust the node relationship weights, regard the power load prediction task as a reinforcement learning problem, and define the agent, environment, state, action and reward. The agent is the weight adjustment module in the parallel graph attention network, the environment is the power load data and its related spatial relationship information, the state is the power load data characteristics and graph structure at the current moment, the action is the adjustment operation of the node relationship weights, and the reward is determined based on the error between the predicted structure and the actual power load.

[0027] In one embodiment of the present invention: the fusion of multi-source heterogeneous data in the expansion of the parallel graph attention network enhances the expression of spatial dependency by fusing meteorological data and geographic information data. The meteorological data and geographic information data are first subjected to feature engineering processing. The meteorological data includes temperature, humidity and wind speed, and the geographic information data includes regional functional division, population density and topography. For meteorological data, the sliding average and wavelet transform methods are used to extract its time series features. For geographic information data, encoding and normalization operations are used to convert it into a numerical feature vector, and the processed multi-source heterogeneous data is fused with the power load data. In the parallel graph attention network, the fused features are used as the input features of the nodes to participate in the generation process of the relationship matrix.

[0028] In one embodiment of the present invention: the data enhancement of the generative adversarial network in the improvement of the joint optimization framework is to enhance the power load data by using the generative adversarial network in the data preprocessing stage. The generative adversarial network consists of a generator and a discriminator. The generator is used to learn the distribution characteristics of the original power load data and generate synthetic data that is the same as the real data. The generator uses a deep convolutional neural network and a variational autoencoder to input a noise vector and output synthetic power load data. The discriminator is responsible for distinguishing between real data and synthetic data, and it uses a multi-layer perceptron structure.

[0029] In one embodiment of the present invention: the adaptive multimodal feature fusion in the improvement of the joint optimization framework adopts an adaptive multimodal feature fusion method in the process of spatiotemporal feature fusion, based on the attention mechanism, automatically adjusts the fusion weights of the modal features according to the data characteristics of the time scale and spatial position, inputs the modal features into their respective feature extraction networks, extracts feature vectors, and calculates the correlation between each feature vector and the target prediction value.

[0030] In one embodiment of the present invention: The Bi-LSTM model uses a bidirectional structure and a gating mechanism to learn the long-term dependencies of the data based on the data information of the previous and next moments in the power load time series, and effectively predicts missing values. In the process of predicting missing values, the Bi-LSTM model deeply mines the time series characteristics of the power load data and captures the trends and patterns in the data. The feature extraction module provides processed multi-scale frequency domain features to the spatial relationship modeling module and the data processing and feature fusion module. The spatial relationship modeling module uses features combined with reinforcement learning and multi-source heterogeneous data to construct a spatial relationship model.

[0031] The Fourier time series graph learning method for cross-cycle power load forecasting includes data processing, feature extraction, spatial relationship modeling, data enhancement, and feature fusion. The specific operation method is as follows: The isolation forest algorithm and DBSCAN are combined to identify and remove outliers in the power load data. For missing values, the Bi-LSTM model is used to predict and fill in the missing values ​​based on its bidirectional structure and gating mechanism to learn the long-term dependency of the data. The Min-Max normalization method is used to normalize the data to Interval, divide the original power load data into daily, weekly and monthly granularity to construct multi-scale time series, use multi-scale Fourier transform, dynamic spectrum reconstruction and multi-resolution spectrum fusion to extract features; The parallel graph attention network is expanded, reinforcement learning is introduced to dynamically adjust the node relationship weights, the prediction task is regarded as a related element of the reinforcement learning problem definition, and multi-source heterogeneous data such as meteorological and geographic information are integrated. After feature engineering processing, they participate in the generation of relationship matrices, build spatial relationship models, and improve the joint optimization framework. The generative adversarial network is used to enhance data in the data preprocessing stage. The generator generates synthetic data, and the discriminator distinguishes between true and false. An adaptive multimodal feature fusion method is adopted. Based on the attention mechanism, the fusion weights of different modal features are adjusted according to the characteristics of time and space data to improve prediction accuracy.

[0032] In one embodiment of the present invention, during dynamic spectrum reconstruction, when abnormal fluctuations in power load are detected within a certain time period, the dynamic model predicts the impact of these fluctuations on the frequency domain characteristics based on learned patterns, and automatically adjusts the amplitude and phase of the frequency domain characteristics. The specific adjustment process is as follows: First, the similarity between the current data and similar fluctuation patterns in historical data is calculated. An adjustment coefficient is then determined based on this similarity. Then, the amplitude and phase of the frequency domain characteristics are fine-tuned according to the adjustment coefficient, ensuring that the frequency domain characteristics more accurately reflect the actual changes in the current power load.

[0033] In one embodiment of the present invention: in multi-resolution spectrum fusion, for the high-frequency features at the daily scale, when fusing them with the medium-frequency features at the weekly scale, the high-frequency features at the daily scale are resampled according to the time intervals of the weekly scale so that they match the medium-frequency features at the weekly scale in terms of time resolution. Then, a weighted average method is used to fuse the two. The weights are determined according to the importance of the features at different time scales. The importance is evaluated based on the feature's ability to explain changes in historical power load data, and is determined by calculating the correlation between the features and the actual load data. The fused composite features can more comprehensively reflect the changing laws of power load at different time scales. Multi-resolution spectrum fusion designs a hierarchical fusion algorithm. For the high-frequency features at the daily scale and the medium-frequency features at the weekly scale, they are first resampled to match the time resolution, and a weighted fusion formula is used: , where the weight Based on the correlation between characteristics and historical power load data changes calculate:

[0034] Similarly, features of other scales are fused to make more accurate predictions of power load by using multi-scale Fourier transform and parallel graph attention network.

[0035] In one embodiment of the present invention: in the introduction of reinforcement learning to dynamically adjust weights, the deep Q network adopted is used as the reinforcement learning algorithm. The intelligent agent queries the corresponding action value function in the Q network according to the current state, and selects the action with the greatest value to execute, that is, adjusts the corresponding node relationship weights. When the prediction error is small, a positive reward is given, otherwise a negative reward is given. Through continuous interactive learning with the environment, the intelligent agent can gradually find the optimal weight adjustment strategy, so that GAT can adapt to the dynamic changes of user behavior and spatial dependencies in the power system more quickly and accurately. When encountering an emergency, the reinforcement learning mechanism can quickly adjust the node relationship weights according to the new state, thereby accurately reflecting the real-time interactive relationship of user events and improving the accuracy of the prediction.

[0036] In one embodiment of the present invention: in the fusion of multi-source heterogeneous data to enhance the expression of spatial dependencies, when calculating the attention weights between nodes, not only the correlation between power load data is considered, but also the environmental factors reflected by meteorological data and geographic information data are taken into consideration. In this way, the model can better explain the differences in electricity consumption in different regions under different environmental conditions, more accurately characterize the spatial dependencies between users, and improve the modeling capabilities of complex spatial scenes.

[0037] In one embodiment of the present invention: in the content of data enhancement of the generative adversarial network, during the training process, the generator and the discriminator perform adversarial training, the generator attempts to generate more realistic synthetic data to deceive the discriminator, and the discriminator continuously improves its ability to distinguish real data from synthetic data. Through this adversarial training, the generator can gradually generate high-quality supplementary data and enrich the diversity of data. Especially for some scenarios where data is scarce, data enhancement technology can effectively expand the amount of data, so that the model can learn more sample features during the training process, thereby improving the model's generalization ability and prediction accuracy.

[0038] For example, please see the attached Figure 1 : Background of the embodiments To verify the effectiveness of the Fourier time series graph learning system for cross-cycle power load forecasting in this paper, we selected power load data from a large city as the test object. This city has a rich electricity consumption scenario, covering various types of electricity consumption, such as industrial, commercial, and residential. It is significantly affected by factors such as seasons, weather, and holidays, and the power load changes are complex. This is consistent with the load characteristics of the current power system and can fully verify the performance of this system in actual applications. Data processing, Outlier processing: We collected 15-minute power load data from the city over the past three years and used the isolation forest algorithm and density-based spatial clustering algorithm to identify outliers. For a particular week's data, we calculated the outlier score for each data point using the isolation forest algorithm. We found that some data points had outlier scores greater than the set threshold of 0.8 and were initially identified as outliers. We then used the DBSCAN algorithm to reconfirm these initially identified outliers and remove any misidentified points. Missing value filling: For data with missing values, an interpolation model based on a bidirectional long short-term memory network (Bi-LSTM) is used to fill them. Assume that there are missing values ​​in the 100th to 105th time steps in the power load data of a certain day. The Bi-LSTM model learns the long-term dependency of the data based on the data information of the previous and next moments through the calculation of the forget gate, input gate, candidate memory unit, memory unit, output gate and hidden state, and predicts and fills the missing values. After calculation by the model, the missing values ​​are predicted to be [specific predicted value 1, specific predicted value 2, ; Data normalization: Min-Max normalization method is used to normalize the processed power load data to Interval, assuming the minimum value in the original data is 500 (megawatts) and the maximum value is 5000 (megawatts), for a data point with a value of 1500 (megawatts), the normalized value is (1500-500) / (5000-500)≈0.22; Feature extraction, Multi-scale time series division: The original power load data is divided into daily, weekly, and monthly granularities. Assuming the total number of time steps is 17,520 (three years of data, one data point every 15 minutes), there are 96 time steps in a day, 672 time steps in a week, and an average of 2,880 time steps in a month, the daily scale series can be divided into 182 (17,520 ÷ 96), the weekly scale series can be divided into 26 (17,520 ÷ 672), and the monthly scale series can be divided into 6 (17,520 ÷ 2,880). Multiscale Fourier transform, Dynamic spectrum reconstruction: Fourier transforms are performed on time series at various scales to obtain frequency domain features. A dynamic model based on gated recurrent units (GRUs) is constructed to monitor data trends. During a certain summer week, the power load experienced abnormal fluctuations. The dynamic model learned this fluctuation pattern by calculating update gates, reset gates, candidate hidden states, and hidden states. The adjustment coefficient for future frequency domain feature changes was predicted based on the hidden states. The designed adjustment coefficient was 1.2, and the corresponding frequency domain feature amplitude and phase were fine-tuned to make the frequency domain features more consistent with actual changes. Multi-resolution spectrum fusion: After Fourier transforming daily, weekly, and monthly data to obtain frequency domain features, a hierarchical fusion algorithm is designed. Taking the fusion of daily high-frequency features and weekly medium-frequency features as an example, the daily high-frequency features are first resampled according to the weekly time interval to match the weekly medium-frequency features in terms of time resolution. Then, the two are fused using a weighted average method. The weights are determined based on the correlation between the features and changes in historical power load data. Calculation shows that the correlation between daily high-frequency features and actual load data is 0.6, and the correlation between weekly medium-frequency features is 0.4. The fusion weights are 0.6 and 0.4, respectively. The fused composite features can more comprehensively reflect the changing patterns of power load at different time scales. Spatial relationship modeling, Introducing reinforcement learning to dynamically adjust weights: A reinforcement learning mechanism is introduced based on the parallel graph attention network (GAT). The power load forecasting task is regarded as a reinforcement learning problem. The agent is defined as the weight adjustment module in the GAT. The environment is the power load data and its related spatial relationship information. The state is the power load data characteristics and graph structure at the current moment. The action is the adjustment operation of the node relationship weight. The reward is determined based on the error between the predicted result and the actual power load. A deep Q network is used as the reinforcement learning algorithm. At a certain moment, the agent queries the action value function in the Q network based on the current state and selects the action with the largest value to adjust the node relationship weight. When the prediction error is small, a positive reward is given, otherwise a negative reward is given. After multiple iterative learning, the agent can find the optimal weight adjustment strategy, enabling the GAT to quickly adapt to the dynamic changes of user behavior and spatial dependencies in the power system. Fusion of multi-source heterogeneous data enhances spatial dependency expression: Meteorological data (temperature, humidity, wind speed) and geographic information data (regional functional division, population density, topography) are integrated. For meteorological data, sliding average and wavelet transform methods are used to extract time series features. A sliding average (window size of 3) is calculated for the temperature data of the past week to obtain a smoothed temperature series. For geographic information data, encoding and normalization operations are used to convert it into numerical feature vectors. The processed multi-source heterogeneous data is then fused with power load data. In GAT, the fused features are used as input features for nodes and participate in the generation of the relationship matrix. When calculating the attention weights between nodes, not only the correlation between power load data but also the environmental factors reflected by meteorological and geographic information data are considered, improving the modeling capabilities of complex spatial scenarios. Data enhancement and feature fusion, Data enhancement using generative adversarial networks: During the data preprocessing phase, power load data is enhanced using a generative adversarial network (GAN). The GAN consists of a generator and a discriminator. The generator uses a deep convolutional neural network and a variational autoencoder to input a noise vector to generate synthetic power load data. The discriminator uses a multi-layer perceptron structure to distinguish between real and synthetic data. During training, the generator attempts to generate more realistic synthetic data to deceive the discriminator, while the discriminator continuously improves its recognition ability. After multiple rounds of training, the generator generates a large amount of high-quality supplementary data, enriching the data diversity. For the power load data of newly built commercial areas, where data is scarce, data augmentation techniques are used to expand the data volume, enabling the model to learn more sample features, improving the model's generalization ability and prediction accuracy. Adaptive multimodal feature fusion: During the spatiotemporal feature fusion process, an adaptive multimodal feature fusion method is used. Based on the attention mechanism, the fusion weights of modal features (time domain, frequency domain, and spatial relationship features) are automatically adjusted according to the data characteristics of time scale and spatial location. Each modal feature is input into its own feature extraction network, feature vectors are extracted, and the correlation between each feature vector and the target prediction value is calculated. During holidays, holiday-related features in the time domain features contribute significantly to the prediction results. The adaptive multimodal feature fusion method automatically increases the weight of these features to more accurately capture the special variation patterns of power load during holidays. The feasibility of the method is demonstrated by example. Improved forecasting accuracy: The system of the present invention was used to forecast the city's power load for the next week and compared with the traditional forecasting method (ARIMA model). The results showed that the mean absolute error (MAE) of the system of the present invention was 50 MW and the root mean square error (RMSE) was 70 MW, while the MAE of the ARIMA model was 120 MW and the RMSE was 150 MW. The forecasting accuracy of the system of the present invention is significantly higher than that of the traditional method, and it can more accurately predict power load changes, providing a reliable basis for power system scheduling. Adaptability to complex scenarios: During a large-scale event in the city, the power load experienced unusual changes. The system, using reinforcement learning to dynamically adjust weights and fuse heterogeneous data from multiple sources, quickly adapted to these changes and accurately predicted power load growth during the event and in the surrounding area. Traditional methods, unable to capture these complex spatiotemporal changes in a timely manner, produced predictions that deviated significantly from actual load. This demonstrates the system's greater adaptability and accuracy in handling complex scenarios. Improved data utilization efficiency: In areas where data is scarce, such as newly built industrial parks, the generative adversarial network data enhancement technology of the present invention expands the amount of data. The model can learn more sample features during training, thereby more accurately predicting the power load in the area. In comparison, the prediction accuracy of traditional methods drops significantly when data is insufficient. This proves that the system of the present invention can effectively improve data utilization efficiency and maintain good prediction performance under various data conditions.

[0039] It can be seen from the above embodiments that the Fourier time series graph learning system for cross-cycle power load forecasting of the present invention has high feasibility and effectiveness in practical applications, and can significantly improve the accuracy and reliability of power load forecasting.

[0040] Although the present invention is disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent variations, and modifications made to the above embodiments in accordance with the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection defined by the claims of the present invention.

Claims

1. A Fourier time series graph learning system for cross-cycle power load forecasting, including a feature extraction module, a spatial relationship modeling module, and a data processing and feature fusion module. The multi-scale Fourier transform belongs to the feature extraction module, the parallel graph attention network expansion belongs to the spatial relationship module, and the joint optimization framework improvement belongs to the data processing and feature fusion module. The multi-scale Fourier transform includes dynamic spectrum reconstruction and multi-resolution spectrum fusion. The parallel graph attention network expansion includes the introduction of reinforcement learning to dynamically adjust weights and the fusion of multi-source heterogeneous data to enhance spatial dependency expression. The joint optimization framework improvement includes data enhancement of the generative adversarial network and adaptive multimodal feature fusion. It is characterized by: The dynamic spectrum reconstruction is based on recursive neural networks and their variants. It predicts future change trends based on fluctuation patterns in historical data, dynamically adjusts frequency domain features and captures dynamic changes in data. It also combines the lonely forest algorithm and density-based spatial clustering algorithm to identify and eliminate outliers.

2. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 1 is characterized in that: The dynamic spectrum reconstruction in the multi-scale Fourier transform is to establish a dynamic model to monitor the change trend of the data in real time after Fourier transforming the power load data to obtain the frequency domain characteristics. The dynamic model is constructed based on a recursive neural network and its variants, and the fluctuation patterns in the historical data are collected and stored in the recursive neural network. The future change trend is predicted based on the historical data. For the power load data point, its outlier score in the lonely forest is calculated. When it is greater than the set threshold, it is determined to be an outlier. An interpolation model based on a bidirectional long short-term memory network is used to set the power load time series. , the missing value position is j, and the Bi-LSTM model is used to predict the missing value: The forward propagation formula of the Bi-LSTM model is: Forget Gate: ; Input Gate: ; Candidate memory cells: ; Memory unit: ; Output gate: ; Hidden state: ; Bidirectional hidden state forward and backward Predict missing values: ； in, as well as is the weight matrix, as well as is the bias vector, is the Sigmoid function, is element-wise multiplication, represents the data input to the Bi-LSTM model at time step t, is the hidden state of the Bi-LSTM model at time step t-1, is the memory cell at time step t-1, and j represents the position where the value is missing in the power load time series; Data normalization uses the Min-Max normalization method, and the specific formula is: ； Power load data Normalized to the interval [0,1], Represents the raw power load data.

3. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 2 is characterized in that: The multi-resolution spectrum fusion in the multi-scale Fourier transform is to perform Fourier transform on the daily, weekly and monthly data respectively to obtain the corresponding frequency domain features, and then design a hierarchical fusion algorithm to cross-fuse the obtained frequency domain features. The original power load data is divided into multi-scale time series, and the daily scale series, weekly scale series and monthly scale series are respectively: ； in, is the total number of time steps, one day is time steps, one week time steps, one month time steps; According to the dynamic spectrum reconstruction, the time series of each scale are transformed by Fourier transform to obtain the frequency domain features. Based on the gated cyclic unit, a dynamic model is constructed to detect the data change trend. The input of the gated cyclic unit is set as the frequency domain feature sequence. The hidden state is but: Update Gate: ; Reset the gate: ; Candidate hidden states: ; Hidden state: ; Survey coefficients for predicting future frequency domain feature changes based on hidden states , the adjusted frequency domain features ,in Depend on Through the fully connected layer, we can get: ; in is an element in the frequency domain feature sequence at time step t, represents the hidden state at time step t-1, Based on the hidden state Survey coefficients obtained through the fully connected layer.

4. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 3 is characterized by: The introduction of reinforcement learning to dynamically adjust weights in the extension of the parallel graph attention network is to introduce a reinforcement learning mechanism on the basis of the parallel graph attention network to dynamically adjust the node relationship weights, regard the power load prediction task as a reinforcement learning problem, and define the agent, environment, state, action and reward. The agent is the weight adjustment module in the parallel graph attention network, the environment is the power load data and its related spatial relationship information, the state is the power load data characteristics and graph structure at the current moment, the action is the adjustment operation of the node relationship weights, and the reward is determined based on the error between the predicted structure and the actual power load.

5. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 4 is characterized in that: The enhanced spatial dependency expression of the fused multi-source heterogeneous data in the extension of the parallel graph attention network is expressed as the fusion of meteorological data and geographic information data. The meteorological data and geographic information data are first subjected to feature engineering processing. The meteorological data includes temperature, humidity and wind speed, and the geographic information data includes regional functional division, population density and topography. For meteorological data, the sliding average and wavelet transform methods are used to extract its time series features. For geographic information data, encoding and normalization operations are used to convert it into a numerical feature vector, and the processed multi-source heterogeneous data is fused with the power load data. In the parallel graph attention network, the fused features are used as the input features of the nodes to participate in the generation process of the relationship matrix.

6. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 5 is characterized in that: The data enhancement of the generative adversarial network in the joint optimization framework is to enhance the power load data by using the generative adversarial network in the data preprocessing stage. The generative adversarial network consists of a generator and a discriminator. The generator is used to learn the distribution characteristics of the original power load data and generate synthetic data that is the same as the real data. The generator uses a deep convolutional neural network and a variational autoencoder to input a noise vector and output synthetic power load data. The discriminator is responsible for distinguishing between real data and synthetic data, and it uses a multi-layer perceptron structure.

7. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 6 is characterized in that: The adaptive multimodal feature fusion in the improvement of the joint optimization framework adopts an adaptive multimodal feature fusion method in the process of spatiotemporal feature fusion. Based on the attention mechanism, the fusion weights of the modal features are automatically adjusted according to the data characteristics of the time scale and spatial position, and the modal features are input into their respective feature extraction networks to extract feature vectors, and the correlation between each feature vector and the target prediction value is calculated.

8. The Fourier time series graph learning system for cross-cycle power load forecasting according to claim 7 is characterized in that: The Bi-LSTM model uses a bidirectional structure and gating mechanism to learn the long-term dependencies of data based on the data information of the previous and next moments in the power load time series, and effectively predicts missing values. In the process of predicting missing values, the Bi-LSTM model deeply mines the time series characteristics of the power load data and captures the trends and patterns in the data. The feature extraction module provides processed multi-scale frequency domain features to the spatial relationship modeling module and the data processing and feature fusion module. The spatial relationship modeling module uses features combined with reinforcement learning and multi-source heterogeneous data to construct a spatial relationship model.

9. A Fourier time series graph learning method for cross-cycle power load forecasting applicable to the Fourier time series graph learning system for cross-cycle power load forecasting according to any one of claims 1 to 8, characterized in that: The specific operation steps of the Fourier time series graph learning method include data processing, feature extraction, spatial relationship modeling, and data enhancement and feature fusion. The specific operation method is as follows: The isolation forest algorithm and DBSCAN are combined to identify and remove outliers in power load data. For missing values, the Bi-LSTM model is used to predict and fill in the missing values ​​based on its bidirectional structure and gating mechanism, learning long-term data dependencies. The Min-Max normalization method is used to normalize the data to the [0, 1] range. The raw power load data is divided into daily, weekly, and monthly granularities to construct a multi-scale time series. Multi-scale Fourier transforms, dynamic spectrum reconstruction, and multi-resolution spectrum fusion are used to extract features. The parallel graph attention network is expanded, reinforcement learning is introduced to dynamically adjust the node relationship weights, the prediction task is regarded as a related element of the reinforcement learning problem definition, and multi-source heterogeneous data such as meteorological and geographic information are integrated. After feature engineering processing, they participate in the generation of relationship matrices, build spatial relationship models, and improve the joint optimization framework. The generative adversarial network is used to enhance data in the data preprocessing stage. The generator generates synthetic data, and the discriminator distinguishes between true and false. An adaptive multimodal feature fusion method is adopted. Based on the attention mechanism, the fusion weights of different modal features are adjusted according to the characteristics of time and space data to improve prediction accuracy.

Images