Generation power partition collaborative prediction method based on space-time correlation modeling and DQN error correction

Abstract

The invention belongs to the technical field of short-term prediction of new energy power of a power system, and provides a generation power partition collaborative prediction method based on space-time correlation modeling and DQN error correction, which comprises the steps of data collection, preprocessing, relation extraction, feature aggregation processing, data modeling processing, correction processing and weighted integration. According to the method, joint modeling is carried out on the spatio-temporal data of the multi-region new energy stations by constructing the spatial topological relation undirected graph, the graph convolution, the time convolution and the Transform combination model, the prediction error is corrected by using the deep Q network algorithm, the spatio-temporal correlation of the new energy power generation data is effectively captured, the adaptive error correction of the prediction result is realized, and the prediction efficiency is improved. The accuracy of short-term partition prediction of the new energy power is improved, and a more reliable decision basis is provided for power grid dispatching and new energy consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of short-term power prediction of new energy power system, in particular to a power generation power zoned collaborative prediction method based on space-time correlation modeling and DQN error correction. BACKGROUND

[0002] In recent years, the field of new energy has developed rapidly, especially in the fields of wind power and photovoltaic power generation. However, at the same time, the contradiction of new energy consumption has become increasingly prominent. The inherent volatility and intermittency of new energy have brought great challenges to the safe operation and consumption of power grids. Accurate and reliable new energy power prediction is the key to solving these problems. Existing new energy power prediction methods can be roughly divided into two categories: one is the traditional method based on physical model and statistical method, which depends on accurate numerical weather prediction and physical parameters of the station. The model requires high accuracy of parameters and is difficult to effectively capture nonlinear relationships. The second is a data-driven method, such as machine learning and deep learning, which can directly learn the power law from historical data and has stronger nonlinear fitting ability. The latter has become the mainstream of current research. With the increasing complexity of power grid dispatching tasks, power generation prediction of a single station has been difficult to meet the needs of power grid dispatching, and regional new energy power prediction has become the focus of current research. The zoned dynamic collaborative prediction requires the model to have the ability to process time and space information.

[0003] Existing technical solutions can only achieve power generation prediction of a single new energy station, lack the ability to process spatial dimension information, and cannot be applied to regional new energy power prediction tasks, which greatly restricts regional new energy consumption. Data-driven methods usually output a fixed prediction value, lack adaptive correction ability for prediction error, and are difficult to dynamically adjust according to prediction deviation, which cannot provide sufficient reliable reference for power grid dispatching and decision-making. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a power generation power zoned collaborative prediction method based on space-time correlation modeling and DQN error correction, which solves the problems of lack of spatial dimension information processing ability and lack of adaptive correction ability for prediction error of existing methods.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A power generation power zoned collaborative prediction method based on space-time correlation modeling and DQN error correction, comprising:

[0007] Collecting historical power data, meteorological feature data, and electrical topology structure and geographical position data of multiple regional new energy stations to obtain original collected data;

[0008] The original collected data is subjected to outlier detection and cleaning by using a DBSCAN algorithm to obtain cleaned data, the cleaned data is subjected to normalization processing to obtain preprocessed data;

[0009] The preprocessed data is subjected to dynamic partition processing and spatial topological relationship extraction to obtain an undirected graph;

[0010] The undirected graph is subjected to feature aggregation processing by using a GCN model to obtain a node feature output matrix;

[0011] The node feature output matrix is subjected to modeling processing by using a TCN-Transformer model to obtain a time-dimension feature sequence;

[0012] The time-dimension feature sequence is input into a DQN agent for correction processing to obtain a correction result;

[0013] All the correction results are summed and integrated to obtain a predicted power value.

[0014] Preferably, the expression of the undirected graph is: ; wherein, ;

[0015] ; is the undirected graph; is a node set; is an edge set; is an original adjacency matrix; is a spatial correlation strength between nodes and ; is an intra-partition correlation enhancement coefficient; is an inter-partition correlation decay coefficient; is a cross-partition correlation threshold; is a normalized correlation strength measure between a field station and a field station ; are respectively a first weight coefficient, a second weight coefficient, and a third weight coefficient; is an electrical distance-based correlation degree; is a geographical distance-based correlation degree; is a historical output Pearson correlation coefficient-based correlation degree.

[0016] Preferably, the node feature output matrix is subjected to modeling processing by using a TCN-Transformer model to obtain a time-dimension feature sequence, including:

[0017] The node feature output matrix is reshaped into a time sequence format;

[0018] calculating the node feature output matrix by using a TCN network to obtain an extended receptive field sequence;

[0019] capturing global long-term dependencies of the extended receptive field sequence by using a Transformer network to obtain encoded features;

[0020] linearly mapping the encoded features by using a fully connected layer to obtain the time dimension feature sequence.

[0021] Preferably, the time dimension feature sequence is input to a DQN agent for correction processing to obtain a correction result, including:

[0022] setting a state space; an expression of the state space is:

[0023] ; wherein, is the state space; is a current relative error; is a first-order difference of the error; is a second-order difference of the error;

[0024] setting an adjustment formula; the adjustment formula is: ; wherein, ; is an adjusted prediction value; is a current preliminary prediction value; is a standardized adjustment ratio; is a maximum allowed adjustment amplitude;

[0025] setting a composite reward function; an expression of the composite reward function is: ; wherein, ; ; ; is the composite reward function; is a core reward item; is an adjustment reward; is a directional reward; is a consistency reward; is a corrected error; is the i-th item in a historical improvement result set;

[0026] constructing the DQN agent according to the state space, the adjustment formula and the composite reward function;

[0027] correcting the time dimension feature sequence by using the DQN agent to obtain the correction result.

[0028] ​Preferably, the GCN model is used for feature aggregation processing on the undirected graph to obtain a node feature output matrix, comprising:

[0029] An input feature matrix is defined, and the expression of the input feature matrix is: ; wherein, is the input feature matrix; is a feature vector of an i-th station; is a total number of stations.

[0030] The GCN model is constructed, and the expression of a convolution layer of the GCN model is:

[0031] ; wherein, ; is an output of an i-th layer convolution layer; is a nonlinear activation function ReLU; is a fused adjacency matrix after self-connection; is a unit matrix; is a degree matrix of ; is a node feature matrix of an i-th layer; is a trainable weight matrix of an i-th layer; The GCN model is used to calculate the input feature matrix to obtain the node feature output matrix. The following technical effects are disclosed in the application:

[0032]

[0033] The application provides a power generation power partition cooperative prediction method based on space-time correlation modeling and DQN error correction.

[0034] The application provides a power generation power partition cooperative prediction method based on space-time correlation modeling and DQN error correction. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor under the premise of not paying creative labor. ​​​

[0036] Figure 1 A schematic diagram of the power generation partition collaborative prediction process based on spatiotemporal correlation modeling and DQN error correction provided in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of GCN-TCN-Transformer provided in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram illustrating the working principle of DQN provided in an embodiment of the present invention;

[0039] Figure 4 A comparison chart of the prediction results of GCN-TCN-Transformer and the prediction results after DQN error correction provided in the embodiments of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a regional collaborative prediction method for power generation based on spatiotemporal correlation modeling and DQN error correction, which solves the problems of existing methods lacking spatial dimension information processing capabilities and lacking adaptive correction capabilities for prediction errors.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a schematic diagram of the power generation zone collaborative prediction process based on spatiotemporal correlation modeling and DQN error correction provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a regional collaborative prediction method for power generation based on spatiotemporal correlation modeling and DQN error correction, comprising:

[0044] Step 100: Collect historical power data, meteorological characteristic data, electrical topology and geographical location data of new energy power stations in multiple regions to obtain raw data;

[0045] Step 200: Use the DBSCAN algorithm to detect and clean outliers in the original collected data to obtain cleaned data. Then, normalize the cleaned data to obtain preprocessed data.

[0046] Step 300: dynamic partition processing and spatial topological relationship extraction are performed on the pretreated data to obtain an undirected graph;

[0047] Step 400: a GCN model is used to perform feature aggregation processing on the undirected graph to obtain a node feature output matrix;

[0048] Step 500: a TCN-Transformer model is used to model process the node feature output matrix to obtain a time-dimension feature sequence;

[0049] Step 600: the time-dimension feature sequence is input into a DQN agent for correction processing to obtain a correction result;

[0050] Step 700: all the correction results are summed and integrated to obtain a predicted power value.

[0051] Specifically, the expression of the undirected graph is: ; wherein, ;

[0052] ; is the undirected graph; is a node set; is an edge set; is an original adjacency matrix; is a spatial correlation strength between nodes and ; is an intra-partition correlation enhancement coefficient; is an inter-partition correlation decay coefficient; is a cross-partition correlation threshold; is a normalized correlation strength measure between a field station and a field station ; are respectively a first weight coefficient, a second weight coefficient, and a third weight coefficient; is an electrical distance-based correlation degree; is a geographical distance-based correlation degree; is a historical output Pearson correlation coefficient-based correlation degree.

[0053] Further, the TCN-Transformer model is used to model process the node feature output matrix to obtain a time-dimension feature sequence, including:

[0054] The node feature output matrix is reshaped into a time sequence format;

[0055] The TCN network is used to calculate the node feature output matrix to obtain an extended receptive field sequence;

[0056] The Transformer network is used to capture global long-term dependencies of the extended receptive field sequence to obtain an encoding feature;

[0057] A full connection layer is used to linearly map the encoding feature to obtain the time dimension feature sequence.

[0058] Specifically, the time dimension feature sequence is input into a DQN agent for correction processing to obtain a correction result, including:

[0059] A state space is set; the expression of the state space is:

[0060] ; wherein, is the state space; is the current relative error; is the first order difference of the error; is the second order difference of the error;

[0061] An adjustment formula is set; the adjustment formula is: ; wherein, ; is the adjusted prediction value; is the current preliminary prediction value; is the standardized adjustment ratio; is the maximum allowed adjustment amplitude;

[0062] A composite reward function is set; the expression of the composite reward function is: ; wherein, ; ; ; is the composite reward function; is a core reward item; is an adjustment reward; is a directional reward; is a consistency reward; is the corrected error; is the i-th item in the historical improvement result set;

[0063] The DQN agent is constructed according to the state space, the adjustment formula, and the composite reward function;

[0064] The time dimension feature sequence is corrected by the DQN agent to obtain the correction result.

[0065] Further, a GCN model is used to perform feature aggregation processing on the undirected graph to obtain a node feature output matrix, including:

[0066] ​An input feature matrix is defined; the expression of the input feature matrix is: ; wherein, is the input feature matrix; is the feature vector of the th station; is the total number of stations.

[0067] The GCN model is constructed; the expression of the convolutional layer of the GCN model is:

[0068] ; wherein, ; is the output of the +1th convolutional layer; is a nonlinear activation function ReLU; is the adjacency matrix after fusion of the self-connection; is an identity matrix; is the degree matrix of ; is the node feature matrix of the th layer; is the trainable weight matrix of the th layer;

[0069] The input feature matrix is calculated by using the GCN model to obtain the node feature output matrix.

[0070] Specifically, the present embodiment takes the new energy stations of multiple prefecture-level cities under the jurisdiction of a certain provincial power grid as an application scenario, and specifically describes a new energy generation short-term power partition collaborative prediction method based on spatio-temporal correlation modeling and DQN error correction. The new energy generation short-term power partition collaborative prediction method based on spatio-temporal correlation modeling and DQN error correction comprises the following steps:

[0071] S1: Collect the historical power data and related meteorological data of the new energy stations in the above-mentioned regions, as well as the geographical position information and electrical topology graph of the stations, and construct an original data set. Remove outliers and fill in missing values from the original data set, normalize all data using the maximum-minimum method, and unify the dimensions of different features.

[0072] An outlier removal is performed by using a density-based spatial clustering algorithm (DBSCAN). Given a neighborhood radius and a minimum point number , the density of points in the neighborhood of each point is judged to identify abnormal points. Points that do not belong to any cluster, i.e. areas with too low density, are considered outliers and are removed from the dataset, thereby obtaining the data after removing outliers. On the basis of removing outliers, linear interpolation is used to fill in missing values in the data after removing outliers. The dataset after the above data cleaning and missing value filling is normalized, and the normalization formula is as follows:

[0073]

[0074] In the formula, is the data before standardization, is the corresponding normalized data, and is the maximum and minimum value in the sequence.

[0075] S2: Referring to Figure 2 , a GCN-TCN-Transformer prediction network is constructed, and the processing flow of data in the model is as follows: first, spatial features are extracted through a multi-level GCN network, and then the output is passed to the subsequent TCN-Transformer network to complete the dependence capture of the feature sequence. Specifically, it includes:

[0076] 1) Multi-level graph structure design:

[0077] The graph convolution (GCN) network is used to cooperatively model the spatial correlation within and between partitions. First, each dynamic partition is regarded as a GCN subgraph, and each subgraph internally contains multiple new energy station nodes. Then, a unified adjacency matrix is constructed to describe the connections within the partition and between the partitions. Finally, the GCN model is trained on the adjacency matrix, and the representation of each node is updated by aggregating the feature information of the neighbor nodes, thereby realizing cooperative modeling across partitions at two levels of within and between partitions.

[0078] 2) Graph structure construction:

[0079] Define an undirected graph containing all new energy stations in the partitions. The elements of the adjacency matrix are constructed as follows:

[0080]

[0081] wherein, is the normalized correlation strength between station and station , and the calculation formula is:

[0082]

[0083] in , , And satisfy As a preferred embodiment of this example, the following is taken: =0.4, =0.2, =0.4, and simultaneously set the internal correlation enhancement coefficient of the partition. Inter-regional correlation attenuation coefficient Association threshold .

[0084] 3) Through the above operations, information from both the internal and external partitions is aggregated. The first layer of the GCN maps the input feature dimension to 64 dimensions, and the second layer outputs 32-dimensional features. The output of the GCN network incorporates new node representations that integrate spatial correlation features. This serves as the input for subsequent TCN-Transformer modules. TCN-Transformer network design:

[0085] The node feature matrix output by the GCN module Remodeling to time series format ,in For historical time steps, This transformation embeds spatial features into the time series framework, providing structured input for subsequent time series modeling. The TCN network consists of multiple layers of causal dilated convolutions stacked together, with its first... Layers of input sequences The operation is defined by the following formula:

[0086]

[0087] Where: ∗ represents the causal dilated convolution operation. For the first The convolution kernel of the layer, For the first The bias vector of the layer.

[0088] The receptive field size R of the causal dilated convolution is determined by the kernel size k and the dilation factor d, and is calculated using the following formula:

[0089]

[0090] As the number of TCN layers increases, the receptive field R of the TCN also increases, but its ability to capture short-term fluctuations decreases. In practical applications, adjustments should be made based on the characteristics of the task. In this example, two TCN layers are selected, the kernel size k=3, and the dilation factor per layer is set to... Increasing.

[0091] The Transformer network receives the output of the TCN network. The system generates query, key, and value vectors through linear transformations. Furthermore, it captures the global long-term dependencies of the sequence through an encoder structure. Its core multi-head self-attention (MHSA) mechanism is calculated as follows: The Transformer encoder receives the output Z of the TCN network and captures the global long-term dependencies through the MHSA mechanism. The calculation process is as follows:

[0092]

[0093]

[0094] in: All are learnable parameter matrices;

[0095]

[0096] Key vector Dimensions The number of attention heads. These are the query, key, and value matrices, respectively.

[0097] The output of the Transformer encoder is linearly mapped through a fully connected layer to obtain the final prediction result. ,in To predict future time steps, For output dimensions.

[0098] The parameter settings in the Transformer network are related to the task requirements. In this embodiment, the Transformer encoder is configured with 2 layers, the decoder with 1 layer, the hidden layer dimension is set to 64, the number of attention heads h=8, the attention factor factor=5, and the key dimension is... Feedforward network dimension The GELU activation function, residual connections, and layer normalization are used to ensure training stability.

[0099] S3: Reference Figure 3 Construct a DQN error correction network:

[0100] The working principle of DQN network error correction is as follows: the correction of prediction error is modeled as a Markov decision process. The agent selects the next action according to the current state space. The reduction of error is used as a reward signal to guide the DQN agent to learn the correct correction strategy. Finally, the corrected prediction result is output through iteration.

[0101] According to the preliminary prediction results of the GCN-TCN-Transformer model, the correction process of the prediction error is modeled as a sequential decision problem in the input DQN network, with the minimization of the prediction error as the reward, and through the continuous interaction between the DQN and the environment, the prediction value is iteratively adjusted. The state space of the DQN agent is designed as a multi-dimensional feature vector containing the dynamic characteristics of the current prediction error, which is mathematically represented as:

[0102]

[0103] 、 、 are the relative error, the first-order difference of the error, and the second-order difference of the error, respectively; is the original prediction error, is the true value, is a very small positive number to prevent the denominator from being zero.

[0104] The action space is defined as a continuous space, and the action is a relative adjustment amount for the current preliminary prediction value The final prediction value after adjustment is calculated by the following formula:

[0105]

[0106] In this example, the coefficient , is adjusted dynamically with the current error.

[0107] The reward function is the key to guiding the DQN to learn the optimal error correction strategy, and the composite reward function designed in this embodiment is as follows:

[0108]

[0109]

[0110]

[0111] wherein, 、 、 、 respectively from the error improvement reward, the adjustment amplitude reward, the direction correctness reward, and the historical action consistency reward, guide the DQN module to learn an effective error correction strategy; is the corrected error; is the normalized adjustment ratio; originates from the historical improvement result set .

[0112] A strategy is adopted to balance the exploration and exploitation of the action space by the DQN agent. The mechanism of the agent selecting an action at state is shown in the following formula:

[0113]

[0114] Wherein, represents the set of all possible actions selected by the agent. refers to the operation actually selected and executed by the agent at time point . is the selection probability, ranging from [0, 1]. is the parameter of the Q network. represents that the agent randomly selects an action from the set of all actions with a probability of , represents that the agent selects the action with the highest value considered by the current Q network with a probability of . .

[0115] Through the above design of the DQN network, the accuracy of the prediction result of the GCN-TCN-Transformer can be further improved.

[0116] The prediction result after the DQN correction is compared with the GCN-TCN-Transformer prediction result before the correction, and the result is shown in Table 1. Figure 4 It can be seen that the embodiment can effectively improve the prediction accuracy.

[0117] Table 1

[0118] MAE RMSE GCN-TCN-Transformer 0.9706 61.2683 90.9607 Add DQN Correction 0.9795 45.2590 75.9141

[0119] The beneficial effects of the present application are as follows:

[0120] The present application constructs a spatial topological relation undirected graph, and a combined model of graph convolution, time convolution and Transformer is used to jointly model the spatio-temporal data of multi-region new energy stations. The deep Q network algorithm is used to correct the prediction error, effectively capturing the spatio-temporal correlation of new energy power generation data, realizing adaptive error correction of the prediction result, and improving the accuracy of short-term partition prediction of new energy power.

[0121] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments.

[0122] The principles and implementations of the present application are described in the specific examples, and the above examples are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A regional collaborative prediction method for power generation based on spatiotemporal correlation modeling and DQN error correction, characterized in that, include: Historical power data, meteorological characteristic data, electrical topology and geographical location data of new energy power stations in multiple regions were collected to obtain raw data. The DBSCAN algorithm is used to detect and clean outliers in the original collected data to obtain cleaned data. The cleaned data is then normalized to obtain preprocessed data. The preprocessed data is dynamically partitioned and spatial topology relationships are extracted to obtain an undirected graph; The undirected graph is subjected to feature aggregation processing using the GCN model to obtain the node feature output matrix. The node feature output matrix is ​​modeled and processed using the TCN-Transformer model to obtain the time-dimensional feature sequence; The time-dimensional feature sequence is input into the DQN agent for correction processing to obtain the correction result; All the correction results are summed and integrated to obtain the predicted power value.

2. The method for regional collaborative prediction of power generation based on spatiotemporal correlation modeling and DQN error correction as described in claim 1, characterized in that, The expression for the undirected graph is: ;in, ; ; This is the undirected graph; A set of nodes; Let it be the set of edges; This is the original adjacency matrix; For nodes and Spatial correlation strength between them; This is the correlation enhancement coefficient within the partition; This refers to the interval correlation attenuation coefficient; The threshold for cross-partition association; For station With station A measure of normalized correlation strength between them; These are the first weighting coefficient, the second weighting coefficient, and the third weighting coefficient, respectively. For correlation based on electrical distance; The degree of association is based on geographical distance; The correlation is based on the Pearson correlation coefficient of historical output.

3. The method for regional collaborative prediction of power generation based on spatiotemporal correlation modeling and DQN error correction as described in claim 1, characterized in that, The node feature output matrix is ​​modeled using the TCN-Transformer model to obtain a time-dimensional feature sequence, including: Reshape the node feature output matrix into a time series format; The extended receptive field sequence is obtained by calculating the node feature output matrix using the TCN network; The global long-term dependencies of the extended receptive field sequences are captured using a Transformer network to obtain encoded features; The encoded features are linearly mapped using a fully connected layer to obtain the time-dimensional feature sequence.

4. The power generation regional collaborative prediction method based on spatiotemporal correlation modeling and DQN error correction according to claim 1, characterized in that, The time-dimensional feature sequence is input into the DQN agent for correction processing to obtain the correction result, including: Define a state space; the expression for the state space is: ;in, This is the state space; This represents the current relative error; The first difference of the error; The second difference of the error; The adjustment formula is set as follows: ;in, ; These are the adjusted forecast values; These are the current preliminary forecast values; Adjust the proportions for standardization; This represents the maximum allowable adjustment range. Define a composite reward function: The expression for the composite reward function is: ;in, ; ; ; The composite reward function; As the core reward item; To adjust the rewards; Directional rewards; As a consistency reward; This is the corrected error; The first in the set of historical improvement results item; The DQN agent is constructed based on the state space, the adjustment formula, and the composite reward function; The DQN agent is used to correct the time-dimensional feature sequence to obtain the correction result.

5. The power generation regional collaborative prediction method based on spatiotemporal correlation modeling and DQN error correction according to claim 2, characterized in that, The undirected graph is subjected to feature aggregation using the GCN model to obtain a node feature output matrix, including: Define the input feature matrix; the expression for the input feature matrix is: ;in, The input feature matrix; For the first Feature vectors of each station; This represents the total number of stations. Construct the GCN model; the expression for the convolutional layer of the GCN model is: ;in, ; For the first +1 convolutional layer output; ReLU is a non-linear activation function. This is the adjacency matrix after merging self-connections; It is the identity matrix; for The degree matrix; For the first The node feature matrix of the layer; For the first Layer-trainable weight matrix; The input feature matrix is ​​calculated using the GCN model to obtain the node feature output matrix.

Images