Method and device for predicting net load of distributed energy system, and electronic equipment

Abstract

The invention discloses a net load prediction method and device for a distributed energy system and electronic equipment. The method relates to the field of power systems, and comprises the steps of obtaining load data of a distributed energy system in a target time period, data of renewable energy sources associated with the distributed energy system in the target time period and weather data of the target time period, and obtaining a target data set based on the obtained data; and inputting the target data set into a target prediction model to obtain target net load data, the target prediction model being used for predicting the net load of the distributed energy system, and the target prediction model comprising a model constructed based on a space-time diagram convolutional network and an attention mechanism. According to the invention, the technical problem of low accuracy of a prediction result when the net load of the distributed energy system is predicted based on a time sequence analysis mode in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to a method, device and electronic equipment for predicting net load of a distributed energy system. Background Art

[0002] The development of distributed energy and the advancement of smart grids have increased uncertainty at both the power supply and load ends. When the busbar in the distribution network supplies both the power load and the output of the new energy device, the difference between the power load and the output of the new energy device constitutes the net load. Since the net load is random, short-term prediction of the net load is crucial to maintaining the stable operation of the power grid. At the same time, the regional net load can be used as a reference for adjusting the output of the units and the spinning reserve capacity in the region.

[0003] Net load forecasting is a key step to ensure the safety and stability of the power system. In order to achieve accurate load forecasting, a variety of forecasting methods are used in related technologies, among which the most typical ones include time series analysis, regression analysis and machine learning. Time series analysis, such as the ARIMA model, focuses on linear modeling of historical load data and can effectively capture trends and seasonal changes in the data. However, it may have limitations in dealing with nonlinear changes and emergencies. First, the ability to handle complexity and nonlinear relationships is limited, especially in nonlinear load fluctuations such as sudden line outages and extreme climates. The forecasting models in related technologies are difficult to achieve the expected accuracy. Secondly, the high dependence of the forecasting methods in related technologies on data quality is another significant defect. Noise, missing values ​​or outliers in the original data often lead to deviations in the forecast results. In addition, all potential external factors (for example, weather factors) have not been fully considered in related technologies, affecting the overall accuracy of the forecast.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present invention provide a method, device and electronic device for predicting the net load of a distributed energy system, so as to at least solve the technical problem in the related art of predicting the net load of a distributed energy system based on time series analysis and the low accuracy of the prediction results.

[0006] According to one aspect of an embodiment of the present invention, a method for predicting the net load of a distributed energy system is provided, comprising: acquiring load data of the distributed energy system in a target time period, data of renewable energy associated with the distributed energy system in the target time period, and weather data in the target time period, and obtaining a target data set based on the acquired data; inputting the target data set into a target prediction model to obtain target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model comprises: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism.

[0007] Furthermore, after the target data set is input into the target prediction model, it includes: mapping the data in the target data set into graph structure data to obtain target graph structure data, wherein the target graph structure data includes: multiple nodes and edges representing the relationship between the multiple nodes, the multiple nodes are used to record the data in the target data set, based on the target graph structure data, using the spatiotemporal graph convolutional network and the attention mechanism to determine the target net load data.

[0008] Furthermore, based on the target graph structure data, the target net load data is determined using the spatiotemporal graph convolutional network and the attention mechanism, including: converting the target graph structure data to obtain a target adjacency matrix, wherein the target adjacency matrix is ​​used to record the data represented by the multiple nodes and the degree of correlation between the multiple nodes; based on the target adjacency matrix, the target net load data is determined using the spatiotemporal graph convolutional network and the attention mechanism.

[0009] Furthermore, based on the target adjacency matrix, the target net load data is determined using the spatiotemporal graph convolutional network and the attention mechanism, including: inputting the target adjacency matrix into the spatiotemporal graph convolutional network in the target prediction model to obtain an output result, wherein the output result is used to record the spatiotemporal relationship of the multiple nodes in the target graph structure data; based on the output result, determining the target net load data using the attention mechanism.

[0010] Furthermore, the target graph structure data is transformed to obtain a target adjacency matrix, including: calculating the correlation between nodes in the target graph structure data to obtain a calculation result, and determining an initial adjacency matrix based on the calculation result; predicting the graph structure data of h time steps based on the initial adjacency matrix to obtain a prediction result, and determining the target adjacency matrix based on the prediction result, wherein h is a positive integer.

[0011] Furthermore, the target prediction model also includes: an output layer, which is a fully connected layer, used to output the target net load data, and based on the output result, the target net load data is determined using the attention mechanism, including: using the attention mechanism to assign weights to the data of each time step in the output result to obtain an assignment result; based on the assignment result, weighted summing the data in the output result to obtain a weighted summing result, and inputting the weighted summing result into the output layer to predict the net load of the distributed energy system to obtain the target net load data.

[0012] Furthermore, the target prediction model is obtained in the following manner: obtaining load data of the distributed energy system in a historical time period, data of renewable energy associated with the distributed energy system in the historical time period, weather data of the historical time period and net load data of the distributed energy system to obtain training samples, wherein the historical time period is before the target time period; performing model training on the initial prediction model based on the training samples, and using a preset evaluation method to evaluate the accuracy of the prediction results of the initial prediction model during the model training process, and obtaining the target prediction model when the evaluation results meet preset conditions, wherein the preset evaluation method includes at least one of the following: mean absolute error and mean absolute percentage error, and the initial prediction model is a target prediction model that has not been model trained.

[0013] According to another aspect of an embodiment of the present invention, a net load prediction device for a distributed energy system is also provided, including: an acquisition unit, used to acquire load data of the distributed energy system in a target time period, data of renewable energy associated with the distributed energy system in the target time period, and weather data in the target time period, and obtain a target data set based on the acquired data; a processing unit, used to input the target data set into a target prediction model to obtain target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism.

[0014] Furthermore, the net load prediction device of the distributed energy system also includes: a mapping unit, which is used to map the data in the target data set into graph structure data after the target data set is input into the target prediction model, so as to obtain target graph structure data, wherein the target graph structure data includes: multiple nodes and edges representing the relationship between the multiple nodes, and the multiple nodes are used to record the data in the target data set; a determination unit, which is used to determine the target net load data based on the target graph structure data by utilizing the spatiotemporal graph convolutional network and the attention mechanism.

[0015] Furthermore, the determination unit includes: a conversion subunit, used to convert the target graph structure data to obtain a target adjacency matrix, wherein the target adjacency matrix is ​​used to record the data represented by the multiple nodes and the degree of correlation between the multiple nodes; a determination subunit, used to determine the target net load data based on the target adjacency matrix using the spatiotemporal graph convolutional network and the attention mechanism.

[0016] Furthermore, the subunit is determined, including: a processing module, used to input the target adjacency matrix into the spatiotemporal graph convolutional network in the target prediction model to obtain an output result, wherein the output result is used to record the spatiotemporal relationship of the multiple nodes in the target graph structure data; a determination module, used to determine the target net load data based on the output result using the attention mechanism.

[0017] Furthermore, the conversion subunit includes: a calculation module, used to calculate the correlation between nodes in the target graph structure data, obtain a calculation result, and determine an initial adjacency matrix based on the calculation result; a prediction module, used to predict the graph structure data of h time steps based on the initial adjacency matrix, obtain a prediction result, and determine the target adjacency matrix based on the prediction result, wherein h is a positive integer.

[0018] Furthermore, the target prediction model also includes: an output layer, which is a fully connected layer, used to output the target net load data, and a determination module, including: an allocation submodule, used to use the attention mechanism to assign weights to the data of each time step in the output result to obtain an allocation result; a processing submodule, used to perform weighted summation on the data in the output result based on the allocation result to obtain a weighted summation result, and input the weighted summation result into the output layer to predict the net load of the distributed energy system to obtain the target net load data.

[0019] Furthermore, the target prediction model is obtained through the following sub-units: an acquisition sub-unit, used to acquire the load data of the distributed energy system in a historical time period, the data of renewable energy associated with the distributed energy system in the historical time period, the weather data of the historical time period and the net load data of the distributed energy system, to obtain a training sample, wherein the historical time period is before the target time period; a training sub-unit, used to perform model training on the initial prediction model based on the training sample, and use a preset evaluation method to evaluate the accuracy of the prediction result of the initial prediction model during the model training process, and obtain the target prediction model when the evaluation result meets the preset conditions, wherein the preset evaluation method includes at least one of the following: mean absolute error and mean absolute percentage error, and the initial prediction model is a target prediction model that has not been model trained.

[0020] According to another aspect of an embodiment of the present invention, there is also provided an electronic device, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute any one of the above-mentioned methods for predicting the net load of a distributed energy system by executing the executable instructions.

[0021] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for predicting the net load of a distributed energy system.

[0022] In the present invention, the load data of the distributed energy system in the target time period, the data of the renewable energy associated with the distributed energy system in the target time period, and the weather data in the target time period are obtained, and a target data set is obtained based on the obtained data; the target data set is input into the target prediction model to obtain the target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism. This solves the technical problem in the related art that the net load of the distributed energy system is predicted based on time series analysis, and the accuracy of the prediction result is low. In the present invention, the target prediction model constructed based on the spatiotemporal graph convolutional network and the attention mechanism predicts the net load of the distributed energy system, and achieves the purpose of accurately extracting the features of the load data and renewable energy data of the distributed energy system in the time dimension and the space dimension to improve the accuracy of the net load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0024] Figure 1 is a flow chart of an optional net load prediction method for a distributed energy system according to an embodiment of the present invention;

[0025] Figure 2 is a schematic diagram of a network structure of an optional target prediction model according to an embodiment of the present invention;

[0026] Figure 3 is a schematic diagram of an optional structure of multi-feature aggregation according to an embodiment of the present invention;

[0027] Figure 4 is a schematic diagram of an optional sequence structure according to an embodiment of the present invention;

[0028] Figure 5 is a schematic diagram of an optional graph structure data according to an embodiment of the present invention;

[0029] Figure 6 is a schematic diagram of an optional net load prediction device for a distributed energy system according to an embodiment of the present invention;

[0030] Figure 7 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0033] It should be noted that the user information (including but not limited to user device information, user personal information, etc.), collected information and data (including but not limited to data used for analysis, stored data, displayed data, electronic medical records, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data comply with the relevant laws, regulations and standards of the relevant regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0034] Embodiment 1

[0035] According to an embodiment of the present invention, a method embodiment of an optional net load forecasting method for a distributed energy system is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] Figure 1 is a flow chart of an optional method for predicting net load of a distributed energy system according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0037] Step S101, acquiring load data of a distributed energy system in a target time period, data of renewable energy sources associated with the distributed energy system in the target time period, and weather data in the target time period, and obtaining a target data set based on the acquired data.

[0038] The above-mentioned target time period can be a period of time before the current moment. The above-mentioned load data includes: the power load of the distributed energy system. The above-mentioned renewable energy data can include: wind power generation and photovoltaic power generation data. The above-mentioned target data set can be composed of the load data of the distributed energy system in the target time period, the data of renewable energy associated with the distributed energy system in the target time period and the weather data in the target time period.

[0039] Step S102, inputting the target data set into the target prediction model to obtain the target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on the spatiotemporal graph convolutional network and the attention mechanism.

[0040] The above-mentioned target prediction model may include: a model built based on a spatiotemporal graph convolutional network and an attention mechanism, so as to fully explore the correlation between the source (power source) and the load (load) in the distributed energy system, and solve the problems of information loss and insufficient model accuracy in related technologies. In this embodiment, STGCN (spatial-temporal graph convolutional network) is used for overall modeling, and the attention mechanism is combined to enhance the model's feature extraction ability for time series data, thereby improving the prediction accuracy and robustness of the target prediction model. The above-mentioned target net load data may include: the data of the net load of the distributed energy system within a period of time after the current moment.

[0041] In this embodiment, the target data set may be input into the target prediction model, and the target prediction model may be used to predict the net load of the distributed energy system to obtain the above-mentioned target net load data.

[0042] Through the above steps, in this embodiment, the target prediction model constructed based on the spatiotemporal graph convolutional network and the attention mechanism predicts the net load of the distributed energy system, and achieves the purpose of accurately extracting the features of the load data and renewable energy data of the distributed energy system in the time dimension and the space dimension to improve the accuracy of the net load prediction. This solves the technical problem of low accuracy of the prediction results when predicting the net load of the distributed energy system based on time series analysis in the related art.

[0043] Figure 2 is a schematic diagram of a network structure of an optional target prediction model according to an embodiment of the present invention. Figure 2 The process of predicting net load using the target prediction model is described below:

[0044] First, graph structure data construction and preprocessing can collect historical data of multiple loads (corresponding to load data), renewable energy output, and weather factors in the integrated energy system (corresponding to the distributed energy system) (i.e., obtain the target data set) and map them into a graph structure (or graph structure data). By calculating the correlation between the variables in the graph structure, an adjacency matrix is ​​constructed to express the nodes in the power network (such as power demand points and renewable energy generation points in different regions) and the edges between nodes (such as the association between loads and energy) as a graph structure according to the geographical location, thereby capturing the spatiotemporal relationship between them.

[0045] Then, a spatiotemporal graph convolutional network (STGCN) is constructed to aggregate and predict the spatiotemporal relationship features. By alternating between graph convolution and temporal convolution, STGCN can effectively capture the complex dependencies in spatiotemporal data. This construction process embodies the aggregation of spatial and temporal features, providing a powerful model capability for spatiotemporal series prediction.

[0046] Finally, the attention mechanism is used to enhance the feature extraction capability of the STGCN model to improve the accuracy and stability of net load prediction. The attention mechanism gives different weights to different features, enabling the target prediction model to more effectively focus on the most critical information for the current prediction task, thereby improving the prediction performance of the target prediction model.

[0047] The network structure of the target prediction model is explained as follows:

[0048] ① Graph structure data construction and preprocessing (time series graph structure data): Collect and organize historical data such as renewable energy output, load, weather factors, etc. in the integrated energy system. Map these data into a graph structure to generate nodes (such as different power demand points and new energy power generation points) and edges (representing the relationship between nodes).

[0049] ② Input layer (input conversion): In the input layer, graph structured data is converted into a format suitable for neural network processing for spatial-temporal embedding.

[0050] ④ Space-time graph convolution layer (STGCN) (space-time embedding): In this layer, the spatiotemporal relationship features between nodes and their neighbors can be extracted through graph convolution operations. Through the graph convolution network (GCN) and the temporal convolution network (TCN), complex spatiotemporal patterns can be captured step by step.

[0051] ⑤Output layer (output mapping): Integrate and transform the features after multi-layer graph convolution operations to generate new feature representations.

[0052] ⑥Attention mechanism layer: The introduction of the attention mechanism allows the model to dynamically adjust weights according to the importance of different features, thereby improving the model's ability to capture important information.

[0053] ⑦ Prediction results: The result generated by the output layer is the predicted value of net power load, which can be used for short-term power demand prediction and subsequent scheduling decisions.

[0054] The network structure of the target prediction model can effectively cope with the challenges of complex spatiotemporal dependencies and multi-source data integration, and improve the accuracy of net load prediction.

[0055] Optionally, after the target data set is input into the target prediction model, the method includes: mapping the data in the target data set into graph structure data to obtain target graph structure data, wherein the target graph structure data includes: multiple nodes and edges representing the relationship between the multiple nodes, the multiple nodes are used to record the data in the target data set, and based on the target graph structure data, the target net load data is determined by utilizing a spatiotemporal graph convolutional network and an attention mechanism.

[0056] In this embodiment, the data in the target data set can be mapped to nodes in the graph structure data, and the relationships between the data in the target data set can be mapped to edges in the graph structure data to obtain the target graph structure data. For example, based on the STGCN time series encoding method, the attributes of the input feature data (i.e., the target data set) can be used as nodes of the graph network (corresponding to the graph structure data) to obtain the target graph structure data.

[0057] In this embodiment, the spatiotemporal graph convolutional network can be used to extract, aggregate and predict the temporal relationship features and spatial relationship features in the target graph structure data, and then the attention mechanism can enhance the feature extraction capability. Finally, the target net load data is obtained through the fully connected layer (or output layer) of the target prediction model.

[0058] STGCN (Spatiotemporal Graph Convolutional Network) is used to effectively capture the spatiotemporal relationship characteristics, and the attention mechanism is combined to enhance the model's feature extraction ability for time series data, thereby improving the prediction accuracy and robustness of the target prediction model.

[0059] Optionally, based on the target graph structure data, the target net load data is determined using a spatiotemporal graph convolutional network and an attention mechanism, including: converting the target graph structure data to obtain a target adjacency matrix, wherein the target adjacency matrix is ​​used to record the data represented by multiple nodes and the degree of correlation between the multiple nodes; based on the target adjacency matrix, the target net load data is determined using a spatiotemporal graph convolutional network and an attention mechanism.

[0060] In this embodiment, a matrix can be used to record the target graph structure data to transform the target graph structure data to obtain a target adjacency matrix, so that the target prediction model can process the target graph structure data and predict the net load of the distributed energy system.

[0061] STGCN (Spatiotemporal Graph Convolutional Network) is used to effectively capture the spatiotemporal relationship characteristics, and the attention mechanism is combined to enhance the model's feature extraction ability for time series data, thereby improving the prediction accuracy and robustness of the target prediction model.

[0062] Optionally, based on the target adjacency matrix, the target net load data is determined using a spatiotemporal graph convolutional network and an attention mechanism, including: inputting the target adjacency matrix into the spatiotemporal graph convolutional network in the target prediction model to obtain an output result, wherein the output result is used to record the spatiotemporal relationship of multiple nodes in the target graph structure data; based on the output result, determining the target net load data using an attention mechanism.

[0063] The construction of the above-mentioned spatiotemporal graph convolutional network (STGCN): The spatiotemporal graph convolutional network (STGCN) can combine the characteristics of the graph convolutional network (GCN) and the temporal convolutional network (TCN) to process graph structured data with temporal characteristics (i.e., the target adjacency matrix corresponding to the target graph structured data). The spatiotemporal graph convolutional network (STGCN) can simultaneously capture the dependencies of data in the spatial and temporal dimensions.

[0064] The following is a detailed introduction to the implementation principle of STGCN:

[0065] (1) Graph Convolutional Network (GCN): Graph Convolutional Network (GCN) is used to process graph structured data. It extracts node features and their spatial relationships by performing convolution operations on the graph structure.

[0066] Input features: The feature matrix of each node is usually represented as X, and the adjacency matrix (which can be represented as A).

[0067] Convolution operation: Through the graph convolution operation, the features of each node are aggregated and transformed to obtain a new node feature representation. Mathematically, graph convolution can be expressed as: Among them, Z is the output feature matrix, A is the adjacency matrix, W is the learnable weight matrix, and σ is the activation function.

[0068] (2) Temporal Convolutional Network (TCN): It can be used to capture dependencies in time series data. Unlike traditional convolutional networks, TCN uses one-dimensional convolution and has the characteristics of dilated convolution and causal convolution, which can effectively capture long-term dependencies.

[0069] Input features: time series data, usually represented as a series of feature matrices (such as X t ).

[0070] Convolution operation: Through the time convolution operation, time series data is processed to extract time patterns and trends. Time convolution can be expressed as: t =TCN(X t ), where Y t is the output of the temporal convolution.

[0071] (3) Spatiotemporal Graph Convolutional Network (STGCN): Combining GCN and TCN, the spatiotemporal graph convolutional network performs convolution operations simultaneously in time and space, achieving comprehensive modeling of spatiotemporal data.

[0072] Spatial graph convolution: At each time step, a graph convolutional network is used to process node features and capture the spatial dependencies between nodes.

[0073] Temporal convolution layer: Performs temporal convolution operations on data of multiple time steps to capture dynamic changes in time.

[0074] Cross-time step convolution: Convolution operations are performed alternately in time and space to achieve synchronous bonding of spatiotemporal features. Specifically, each layer of STGCN includes one or more graph convolution blocks and temporal convolution blocks, which are stacked alternately to form a deep network structure.

[0075] (4) Overall structure of STGCN,The overall structure of STGCN can be expressed as:

[0076] Input layer: Input node feature matrix and adjacency matrix (corresponding to the target adjacency matrix), as well as time series data.

[0077] Spatial graph convolution layer: applies graph convolution operations to extract spatial dependencies.

[0078] Temporal convolution layer: applies temporal convolution operation to extract temporal dependencies.

[0079] Output layer: The final output layer generates the prediction result Y, where Y = f(Y (T) ,W (o) ), f(·) represents the mapping function of the output layer (such as the fully connected layer).

[0080] The spatiotemporal graph convolutional network (STGCN) can effectively capture the complex dependencies in spatiotemporal data and is suitable for application scenarios such as traffic flow prediction and social network analysis that need to consider spatiotemporal characteristics at the same time. Its core lies in the comprehensive modeling of spatiotemporal features through alternating stacked graph convolution and temporal convolution layers.

[0081] Optionally, the target graph structure data is transformed to obtain a target adjacency matrix, including: calculating the correlation between nodes in the target graph structure data, obtaining a calculation result, and determining an initial adjacency matrix based on the calculation result; predicting the graph structure data of h time steps based on the initial adjacency matrix, obtaining a prediction result, and determining the target adjacency matrix based on the prediction result, wherein h is a positive integer.

[0082] For example, (1) data collection and aggregation, in an embodiment, the attributes of the input feature data (target data set) can be used as nodes of the graph network based on the time series encoding method of STGCN. STGCN can improve the prediction accuracy by extracting the correlation between node features and aggregating node features. Figure 3 is a schematic diagram of an optional multi-feature aggregation structure according to an embodiment of the present invention, wherein the graph structure of the net load prediction and the structure of other inputs are as follows: Figure 3 As shown. First, through multi-feature aggregation Figure 3 As shown in Figure 1, data from different sources (such as wind power generation, photovoltaic power generation, actual load and weather environment, etc.) are integrated to provide a comprehensive information basis for the target prediction model. Then, the sequence structure can be used as Figure 4 As shown in Figure 2, these integrated features are arranged according to time steps to form a time series framework to capture the dynamic associations and patterns between multiple variables. This method of combining multivariate feature aggregation with time series organization enables STGCN to find the optimal expression between the time dimension and multivariate features, thereby significantly improving the accuracy and reliability of net load forecasting.

[0083] (2) Graph structure data construction: In order to build a graph structure, node sets and edge sets are defined to better understand and represent the complex relationships between multiple variables, thereby achieving more accurate predictions. Figure 5 is a schematic diagram of an optional graph structure data according to an embodiment of the present invention, such as Figure 5As shown, the elements in the node set are represented by vi and vj, which represent the historical characteristics of wind power, photovoltaic power, actual load and weather environment factors involved in net load forecasting. The nodes are connected to each other through edges, and the elements in the edge set can be represented by εij.

[0084] The adjacency matrix G can be determined by the topological structure of the graph structure, and its size is N×N. In this embodiment, the maximum information coefficient (MIC) can be used to calculate the correlation between nodes, so as to obtain the weighted edges and adjacency matrix G of the graph. The basic principle of MIC is that if there is a certain relationship between two variables, their scatter plot can be divided into multiple grids, and the relationship between the variables is encapsulated in these grids. The larger the MIC value, the stronger the correlation between the variables. In this way, the relationship between wind power, photovoltaic power generation, actual load and weather environment can be effectively quantified, so as to construct an accurate graph structure.

[0085] For example, after collecting all relevant multivariate time series data (such as historical power load, wind power generation, solar power generation, and weather and environmental parameters, etc.), the target data set is obtained. After constructing the graph structure, the MIC value can be calculated: the MIC value is calculated for each pair of variables (nodes). The higher the MIC value, the stronger the correlation between the two variables, whether it is a linear or nonlinear relationship. The calculation of MIC is based on dividing the joint distribution of the two variables into grids and evaluating the consistency and diversity of the distribution of the variables in these grids. Specifically, MIC finds a gridding method that can maximize the capture of the relationship between the two variables, and then calculates a standardized indicator that reflects the strength of this relationship. Then, a weighted edge and adjacency matrix can be constructed. Specifically, based on the calculated MIC value, the adjacency matrix of the graph can be constructed. In this matrix, each row and column can correspond to a node, and each element in the matrix represents the correlation (MIC value) between the corresponding nodes. These values ​​can be used as edge weights to form weighted edges. In graph neural networks, the adjacency matrix can reflect the connection relationship between nodes, and the edge weights reflect the connection strength, that is, the correlation between variables.

[0086] Graph neural network initialization: With weighted edges and adjacency matrices, graph neural networks can be initialized. In this network, nodes represent variables, and edge weights represent the correlation between variables. Through operations such as graph convolution, the network can learn the complex interactions and influences of these variables in time series, thereby improving the accuracy of predictions.

[0087] On this basis, x (any real number) can be defined as the graph signal of the predictor variable on the adjacency matrix G (corresponding to the initial adjacency matrix), that is, the value of each node on the graph. By training the model function h(·), the m graph signals at the past moment can be mapped to the graph signals of the next 1 to h steps, and the target adjacency matrix including the graph signals of the next 1 to h steps can be obtained. The specific formula is:

[0088] [x t-m ,...,x t-1 ; G]h(·)=[x t ,…,x t+h-1 ]

[0089] Among them, x t represents the graph signal at time t; h represents the number of graph signals in the future; m represents the number of graph signals in the past, G is the adjacency matrix; and h(·) is the function used to train the model.

[0090] Optionally, the target prediction model also includes: an output layer, which is a fully connected layer, used to output target net load data. Based on the output result, the target net load data is determined using an attention mechanism, including: using the attention mechanism to assign weights to the data of each time step in the output result to obtain an allocation result; based on the allocation result, weighted summing the data in the output result to obtain a weighted summing result, and inputting the weighted summing result into the output layer to predict the net load of the distributed energy system to obtain the target net load data.

[0091] In the spatiotemporal graph convolutional network, in order to further improve the accuracy of the model's prediction results, an attention mechanism can be introduced on the output results to form a STCGN-Attention model (corresponding to the target prediction model). The attention mechanism can help the model automatically learn the importance of features of different nodes or time steps in the final prediction, thereby dynamically adjusting their contribution weights.

[0092] The following are the specific steps and formulas to explain how to apply the attention mechanism to the output results of STGCN.

[0093] (1) The purpose of introducing the attention mechanism: In STGCN, the features of different nodes or time steps may contribute differently to the final prediction. For example, in net load forecasting, certain meteorological factors have a greater impact on the results. Through the attention mechanism, the target prediction model can automatically learn these importance weights and dynamically adjust their contributions according to different nodes or time steps.

[0094] (2) Calculation of attention weights: Assume that there is an output Y∈R of a spatiotemporal graph convolutional network N×F×T, where: N is the number of nodes, F is the feature dimension, and T is the number of time steps.

[0095] The goal of the attention mechanism is to assign a weight to each node or time step, indicating its importance in the final prediction.

[0096] ① Node-level attention: At the node level, different weights can be assigned to the features of each node. First, the attention score of each node feature can be calculated:

[0097] e i =AttentionScore(Y i )=v T ·tanh(W·Y i +b)

[0098] Among them, Y i is the feature matrix of node i, W and b are learnable weights and biases, v T is the attention vector, tanh is the hyperbolic tangent activation function, AttentionScore(Y i ) is node Y i Attention score.

[0099] Then, the scores are converted into weights through the softmax function (normalized exponential function):

[0100]

[0101] Among them, α i is the attention weight of node i, e i represents the attention score of node i, e j represents the attention score at time step j.

[0102] ②Time-level attention: Similarly, an attention mechanism can be introduced in the time dimension to assign weights to the features of each time step. Assume that the attention weight is calculated for each time step t:

[0103] e t =AttentionScore(Y .,t )=v T ·tanh(W·Y ·,t +b)

[0104] Among them, Y ·,t is the feature matrix of all nodes at time step t. Then the weight of the time step can also be calculated by the softmax function:

[0105]

[0106] ③Weighted summation: get the attention weight αi or α t After that, the original features can be combined with these weights by weighted summation to get the final representation of node-level attention:

[0107]

[0108] ④Final prediction: attention-weighted feature Y attn It can be used as the final feature representation and passed to the output layer for prediction. Assume that the output layer is a fully connected layer to generate the final prediction result.

[0109]

[0110] Where W (o) 、b (o) are the learnable weights and biases of the output layer.

[0111] Optionally, the target prediction model is obtained in the following manner: obtaining load data of the distributed energy system in a historical time period, data of renewable energy associated with the distributed energy system in the historical time period, weather data of the historical time period and net load data of the distributed energy system to obtain training samples, wherein the historical time period is before the target time period; performing model training on the initial prediction model based on the training samples, and using a preset evaluation method to evaluate the accuracy of the prediction results of the initial prediction model during the model training process, and obtaining the target prediction model when the evaluation results meet preset conditions, wherein the preset evaluation method includes at least one of the following: mean absolute error and mean absolute percentage error, and the initial prediction model is a target prediction model that has not been model trained.

[0112] In this embodiment, historical load data and meteorological data (or weather data) of a historical time period can be collected, and load data and meteorological data collected at 24 points every day in the historical time period with a sampling interval of 1 hour can be obtained. The training features may include power load, wind power generation, and photovoltaic power generation. The training label may be the net load of the power grid (corresponding to the weather data of the historical time period and the net load data of the distributed energy system), and the data set is divided into a training set, a validation set, and a test set according to a preset ratio (for example, a ratio of 7:1:2).

[0113] In order to facilitate the evaluation of the prediction performance of the initial prediction model, the mean absolute error (MAE) and the mean absolute percentage error (MAPE) can be used to evaluate the prediction results. When the mean absolute error (MAE) and the mean absolute percentage error (MAPE) are within the preset range (corresponding to the preset conditions), the initial prediction model can be determined as the target prediction model. Both are common indicators for measuring the accuracy of the predicted value and the actual value. The calculation method is shown in the following formula:

[0114]

[0115] Where N is the length of the test set; is the predicted value; is the true value.

[0116] It should be noted that in order to verify the prediction performance of the STGCN-Attention model (target prediction model) in the present invention, the autoregressive integrated moving average model (ARIMA), recurrent neural network (RNN), long short-term memory network (LSTM), Transformer model and Informer model were selected as comparison objects, and the application effect of the target prediction model in power load forecasting has been explored. When using power load, wind power generation, and photovoltaic power generation as input features in the training set; the net load of the power grid is used as the training label, that is, the prediction target. In order to ensure a fair comparison, each data set was normalized to zero mean, and the specific parameter settings of each model were adjusted according to the characteristics of the load forecasting task.

[0117] The prediction effect of each model is shown in Table 1. In the experiment, the length of the selected input sequence is 168h (7 days) and the length of the prediction sequence is 24h (1 day).

[0118] Table 1

[0119]

[0120]

[0121] The MAE value of STGCN-Attention on the dataset is 0.416, which is 51.46%, 42.62%, 39.09%, 22.68% and 12.26% lower than the average values ​​of ARIMA, RNN, LSTM, Transformer and Informer, respectively. The MAPE values ​​of STGCN-Attention are 2.025%, which is lower than the MAPE values ​​of RNN, LSTM, Transformer and Informer. The results show that the predicted values ​​of RNN, LSTM, Transformer and STGCN-Attention are highly correlated with the actual values. The prediction results of ARIMA are generally stable, and its prediction ability is mainly to track the general trend of the time series. The characteristics of the data itself have little effect on the prediction performance.

[0122] The results show that the error between the prediction results of STGCN-Attention and the actual value is the smallest. The ARIMA model has poor prediction ability and can only track the overall trend of the net load sequence, but cannot handle the drastic fluctuations of the sequence. The LSTM model can predict the results based on the net load sequence very well, mainly because the LSTM can record the net load sequence through hidden states. The Transformer model extracts information of different time periods of the net load sequence through the self-attention mechanism, thereby predicting the overall trend and local information of the net load sequence. Although the correlation is extracted, in the case of some extreme values, there is a certain deviation between the prediction results and the actual values, resulting in a slight decrease in the prediction results. The target prediction model proposed in the present invention can well suppress the prediction error of extreme values, and use STGCN-Attention to perform feature aggregation on multiple features of the load sequence, combined with information such as photovoltaic, wind energy and meteorological environment, to accurately predict the future state of the net load, thereby achieving better prediction results.

[0123] The experimental results show that compared with other data-driven methods (or other net load prediction methods), the prediction model generated by training in the present invention has higher prediction accuracy. At the same time, the net load prediction method proposed in this embodiment can be applied to data of different scales and different prediction lengths, and has strong generalization ability. It should be noted that, on the basis of this embodiment, factors such as demand response potential, distributed energy risk characteristics, and grid state variables related to net load can also be considered to further improve the model's prediction performance for net load data. In addition, in addition to improving the prediction accuracy, it can also be applied from the perspectives of insufficient measurement configuration and the impact of real-time measurement noise.

[0124] In this embodiment, first, the time series of multiple loads, renewable energy output and weather factors in the integrated energy system are mapped to graph structure data, the maximum information coefficient is used to calculate the correlation between variables, the adjacency matrix is ​​constructed, and the graph network is initialized. Secondly, the spatiotemporal graph convolutional network (STGCN) is introduced to extract the spatiotemporal correlation between the source and load of distributed energy. Finally, the attention mechanism is used to enhance the feature extraction ability of STGCN for time series data and improve the accuracy of short-term net load prediction.

[0125] Embodiment 2

[0126] The second embodiment of the present invention provides an optional net load prediction device for a distributed energy system, and each implementation unit in the net load prediction device corresponds to each implementation step in the first embodiment.

[0127] Figure 6 is a schematic diagram of an optional distributed energy system net load prediction device according to an embodiment of the present invention, such as Figure 6As shown, it includes: an acquisition unit 61 and a processing unit 62.

[0128] The acquisition unit 61 is used to acquire the load data of the distributed energy system in the target time period, the data of the renewable energy associated with the distributed energy system in the target time period and the weather data in the target time period, and obtain the target data set based on the acquired data;

[0129] The processing unit 62 is used to input the target data set into the target prediction model to obtain the target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism.

[0130] In the net load prediction device of the distributed energy system provided in the second embodiment of the present invention, the load data of the distributed energy system in the target time period, the data of the renewable energy associated with the distributed energy system in the target time period and the weather data in the target time period can be obtained by the acquisition unit 61, and the target data set is obtained based on the acquired data. The target data set is input into the target prediction model through the processing unit 62 to obtain the target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism. This solves the technical problem of low accuracy of the prediction result in predicting the net load of the distributed energy system based on time series analysis in the related art. In this embodiment, the target prediction model constructed based on the spatiotemporal graph convolutional network and the attention mechanism predicts the net load of the distributed energy system, and achieves the purpose of accurately extracting the features of the load data and renewable energy data of the distributed energy system in the time dimension and the space dimension to improve the accuracy of the net load prediction.

[0131] Optionally, in the net load prediction device of a distributed energy system provided in Example 2 of the present invention, the net load prediction device of the distributed energy system also includes: a mapping unit, which is used to map the data in the target data set into graph structure data after the target data set is input into the target prediction model, so as to obtain target graph structure data, wherein the target graph structure data includes: multiple nodes and edges representing the relationship between the multiple nodes, and the multiple nodes are used to record the data in the target data set; a determination unit, which is used to determine the target net load data based on the target graph structure data by using a spatiotemporal graph convolutional network and an attention mechanism.

[0132] Optionally, in the net load prediction device of the distributed energy system provided in Example 2 of the present invention, the determination unit includes: a conversion subunit, used to convert the target graph structure data to obtain a target adjacency matrix, wherein the target adjacency matrix is ​​used to record the data represented by multiple nodes and the degree of correlation between the multiple nodes; a determination subunit, used to determine the target net load data based on the target adjacency matrix by using a spatiotemporal graph convolutional network and an attention mechanism.

[0133] Optionally, in the net load prediction device of the distributed energy system provided in Example 2 of the present invention, a subunit is determined, including: a processing module, used to input the target adjacency matrix into the spatiotemporal graph convolutional network in the target prediction model to obtain an output result, wherein the output result is used to record the spatiotemporal relationship of multiple nodes in the target graph structure data; a determination module, used to determine the target net load data based on the output result using an attention mechanism.

[0134] Optionally, in the net load prediction device of the distributed energy system provided in Example 2 of the present invention, the conversion subunit includes: a calculation module, used to calculate the correlation between nodes in the target graph structure data, obtain the calculation results, and determine the initial adjacency matrix based on the calculation results; a prediction module, used to predict the graph structure data of h time steps based on the initial adjacency matrix, obtain the prediction results, and determine the target adjacency matrix based on the prediction results, where h is a positive integer.

[0135] Optionally, in the net load prediction device of a distributed energy system provided in Example 2 of the present invention, the target prediction model also includes: an output layer, which is a fully connected layer, used to output target net load data, and a determination module, including: an allocation submodule, used to use an attention mechanism to assign weights to the data of each time step in the output result to obtain an allocation result; a processing submodule, used to perform weighted summation on the data in the output result based on the allocation result to obtain a weighted summation result, and input the weighted summation result into the output layer to predict the net load of the distributed energy system to obtain the target net load data.

[0136] Optionally, in the net load prediction device of a distributed energy system provided in Example 2 of the present invention, the target prediction model is obtained through the following sub-units: an acquisition sub-unit, used to acquire the load data of the distributed energy system in a historical time period, the data of renewable energy associated with the distributed energy system in the historical time period, the weather data of the historical time period and the net load data of the distributed energy system, to obtain training samples, wherein the historical time period is before the target time period; a training sub-unit, used to perform model training on the initial prediction model based on the training samples, and use a preset evaluation method to evaluate the accuracy of the prediction results of the initial prediction model during the model training process, and obtain the target prediction model when the evaluation results meet the preset conditions, wherein the preset evaluation method includes at least one of the following: mean absolute error and mean absolute percentage error, and the initial prediction model is a target prediction model that has not been model trained.

[0137] The above-mentioned net load prediction device of the distributed energy system may also include a processor and a memory. The above-mentioned acquisition unit 61 and processing unit 62 are stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize corresponding functions.

[0138] The above processor includes a kernel, which calls the corresponding program unit from the memory. One or more kernels can be set, and the target prediction model constructed based on the spatiotemporal graph convolutional network and attention mechanism is adjusted to predict the net load of the distributed energy system, so as to accurately extract the characteristics of the load data and renewable energy data of the distributed energy system in the time dimension and the space dimension, so as to improve the accuracy of the net load prediction.

[0139] The above-mentioned memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one storage chip.

[0140] According to another aspect of an embodiment of the present invention, there is also provided an electronic device, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute any one of the above-mentioned methods for predicting the net load of a distributed energy system by executing the executable instructions.

[0141] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for predicting the net load of a distributed energy system.

[0142] Figure 7is a schematic diagram of an electronic device according to an embodiment of the present invention, such as Figure 7 As shown, an embodiment of the present invention provides an electronic device 70, which includes a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, any one of the above-mentioned methods for predicting the net load of a distributed energy system is implemented.

[0143] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0144] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0145] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0146] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0147] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0148] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0149] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting net load of a distributed energy system, characterized in that: include: Acquire load data of a distributed energy system in a target time period, data of renewable energy sources associated with the distributed energy system in the target time period, and weather data of the target time period, and obtain a target data set based on the acquired data; The target data set is input into a target prediction model to obtain target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism.

2. The net load prediction method according to claim 1, characterized in that: After the target data set is input into the target prediction model, the method includes: Mapping the data in the target data set into graph structure data to obtain target graph structure data, wherein the target graph structure data includes: a plurality of nodes and edges representing the relationship between the plurality of nodes, the plurality of nodes being used to record the data in the target data set, Based on the target graph structure data, the target net load data is determined using the spatiotemporal graph convolutional network and the attention mechanism.

3. The net load prediction method according to claim 2, characterized in that: Based on the target graph structure data, the target payload data is determined by using the spatiotemporal graph convolutional network and the attention mechanism, including: Converting the target graph structure data to obtain a target adjacency matrix, wherein the target adjacency matrix is ​​used to record the data represented by the multiple nodes and the correlation between the multiple nodes; Based on the target adjacency matrix, the target net load data is determined using the spatiotemporal graph convolutional network and the attention mechanism.

4. The net load prediction method according to claim 3, characterized in that: Based on the target adjacency matrix, the target payload data is determined using the spatiotemporal graph convolutional network and the attention mechanism, including: Inputting the target adjacency matrix into the spatiotemporal graph convolutional network in the target prediction model to obtain an output result, wherein the output result is used to record the spatiotemporal relationship of the multiple nodes in the target graph structure data; Based on the output result, the target payload data is determined using the attention mechanism.

5. The net load prediction method according to claim 3, characterized in that: The target graph structure data is converted to obtain a target adjacency matrix, including: Calculating the correlation between nodes in the target graph structure data to obtain a calculation result, and determining an initial adjacency matrix according to the calculation result; The graph structure data of h time steps are predicted based on the initial adjacency matrix to obtain a prediction result, and the target adjacency matrix is ​​determined based on the prediction result, wherein h is a positive integer.

6. The net load prediction method according to claim 4, characterized in that: The target prediction model further includes: an output layer, which is a fully connected layer, and is used to output the target net load data. Based on the output result, the target net load data is determined using the attention mechanism, including: Using the attention mechanism to assign weights to the data of each time step in the output result, to obtain an assignment result; Based on the allocation result, the data in the output result is weightedly summed to obtain a weighted summation result, and the weighted summation result is input into the output layer to predict the net load of the distributed energy system to obtain the target net load data.

7. The net load prediction method according to claim 3, characterized in that: The target prediction model is obtained by: Obtaining load data of the distributed energy system in a historical time period, data of renewable energy sources associated with the distributed energy system in the historical time period, weather data of the historical time period, and net load data of the distributed energy system to obtain training samples, wherein the historical time period is before the target time period; The initial prediction model is trained based on the training samples, and the accuracy of the prediction results of the initial prediction model is evaluated by a preset evaluation method during the model training process. When the evaluation results meet the preset conditions, the target prediction model is obtained, wherein the preset evaluation method includes at least one of the following: mean absolute error and mean absolute percentage error, and the initial prediction model is a target prediction model that has not been trained.

8. A net load prediction device for a distributed energy system, characterized in that: include: An acquisition unit, configured to acquire load data of a distributed energy system in a target time period, data of renewable energy sources associated with the distributed energy system in the target time period, and weather data in the target time period, and obtain a target data set based on the acquired data; A processing unit is used to input the target data set into a target prediction model to obtain target net load data, wherein the target prediction model is used to predict the net load of the distributed energy system, and the target prediction model includes: a model constructed based on a spatiotemporal graph convolutional network and an attention mechanism.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the net load prediction method for a distributed energy system according to any one of claims 1 to 7.

10. An electronic device, characterized in that: It includes one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the net load prediction method for a distributed energy system as described in any one of claims 1 to 7.

Images