Distributed photovoltaic power generation power prediction method based on urban micrometeorological model
Through embedded graph convolution-long and short-term memory network based on urban micrometeorological model, the problem of low prediction accuracy of distributed photovoltaic power generation is solved, and efficient power system operation and grid scheduling is achieved.
Patent Information
- Application Number
- CN202410906777.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-08
- Publication Date
- 2024-11-19
AI Technical Summary
The power prediction accuracy of distributed photovoltaic power generation is not high, the time-space dependence processing is insufficient, and the data type is single, making it difficult to meet the grid scheduling needs.
Using an embedded graph convolution-long and short-term memory network (GC-LSTM) structure based on urban micrometeorological model, integrating meteorological data and photovoltaic power generation data, to build urban micrometeorological models, and capture them through graph convolution neural network and long-term short-term memory network The geographical location and time dependence between power stations realize accurate prediction of distributed photovoltaic power generation power.
The accuracy of distributed photovoltaic power generation power prediction and the operating efficiency of the power system are improved, the grid load distribution and energy management are optimized, and the safe, stable and efficient operation of the power system is ensured.
Smart Images

Figure BDA0004932350960000031 
Figure BDA0004932350960000041 
Figure BDA0004932350960000042
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of photovoltaic power generation control, and in particular relates to a distributed photovoltaic power generation power prediction method based on an urban micro-meteorological model. Background Art
[0002] With the deepening of the concept of sustainable development and the continuous advancement of photovoltaic power generation technology, the proportion of photovoltaic power generation in the integrated energy system of the power grid is increasing. However, the volatility and uncertainty of photovoltaic power generation will affect the safe and stable operation of the power system. With the gradual expansion of photovoltaic power generation, it has brought severe challenges to the economy and safety of the power system. Therefore, accurate prediction of photovoltaic power generation is crucial for the safe and stable operation of the power grid. It can estimate the changes in photovoltaic output in advance and help the dispatching department to develop reasonable and efficient solutions.
[0003] The development of photovoltaic technology has promoted the rise of distributed photovoltaic power stations in cities. Taking Hangzhou as an example, as of 2023, the city has completed and is building 67 distributed photovoltaic projects, ranking first in the country. However, due to the relatively small installed capacity and wide distribution of distributed power stations, and the limitations of cost, technical conditions and other factors, these projects lack independent and complete power station monitoring systems, making it difficult to obtain accurate meteorological data. When large-scale distributed photovoltaic power stations are connected to the power grid, the traditional single-station power prediction scheme can no longer meet the dispatching needs of the power system, and it is urgent to carry out regional distributed photovoltaic power prediction research.
[0004] Photovoltaic power generation is greatly affected by weather, and weather changes are a nonlinear and random process. Various weather factors such as light, temperature, and humidity have a significant impact on photovoltaic power generation. For urban distributed photovoltaic power generation clusters, building an urban micro-meteorological model can help improve the prediction accuracy of distributed photovoltaic power generation.
[0005] Therefore, in order to ensure the economy and safety of distributed photovoltaic systems, it is necessary to develop a distributed photovoltaic power generation power prediction method based on urban micrometeorological model to solve the above problems. Summary of the invention
[0006] Aiming at the problems of low prediction accuracy, insufficient processing of spatiotemporal dependency and single data type in the prior art of distributed photovoltaic power generation, the present invention provides a distributed photovoltaic power generation prediction method based on an urban micro-meteorological model. By integrating meteorological data and photovoltaic power generation data, an urban micro-meteorological model is constructed, and an embedded graph convolution-long short-term memory network (GC-LSTM) structure is adopted to achieve accurate prediction of distributed photovoltaic power generation, thereby optimizing the operating efficiency and reliability of the power system.
[0007] A distributed photovoltaic power generation prediction method based on an urban micro-meteorological model comprises: using a constructed embedded graph convolution-long short-term memory network, taking recently acquired meteorological data and power generation data as input, to achieve prediction of power generation for a period of time in the future; the embedded graph convolution-long short-term memory network is composed of a graph convolution neural network and a long short-term memory network, and the output of the graph convolution neural network is used as the input of the long short-term memory network to achieve prediction of the distributed photovoltaic power generation.
[0008] Preferably, in the embedded graph convolution-long short-term memory network, feature graph information including geographic location, meteorological data and power generation data is constructed by a graph convolutional neural network; and the long short-term memory network utilizes the feature graph information and the time series data dependency to obtain the predicted value of the distributed photovoltaic power generation power.
[0009] The present invention mainly includes the steps of data collection and preprocessing, urban micro-meteorological model construction, graph structure design, time dynamic modeling, model training and prediction. The collected meteorological data and photovoltaic power generation data are processed by using a graph convolutional neural network (GCN) to capture the geographical location and local meteorological correlation between power stations, and the long short-term memory network (LSTM) is combined for time dynamic modeling to analyze the changing rules of data in the time dimension.
[0010] The specific contents include the following:
[0011] (1) Data collection and preprocessing
[0012] Meteorological sensors are installed at photovoltaic power stations in different areas of the city to collect meteorological data including temperature, humidity, wind speed and irradiance. At the same time, historical (for example, the past 1 to 5 months) power generation data (i.e. power generation data) of relevant photovoltaic power stations are obtained. Then, all the collected data are cleaned and standardized to provide a reliable data basis for subsequent power prediction research.
[0013] When actually making predictions, the data used are the recent stored meteorological data and photovoltaic power generation data related to the current time. For example, when predicting the current power generation, data from the past 3 to 5 days can be selected, and the same cleaning and standardization processing can be performed before making predictions.
[0014] (2) Construction of urban micro-meteorological model
[0015] A micro-meteorological model, namely an embedded graph convolution-long short-term memory network, is constructed using meteorological data and photovoltaic power generation data. The pre-processed data is processed using a graph convolutional neural network (GCN) to capture the correlation pattern between the geographical locations of distributed power stations and local meteorological conditions. In this way, a more accurate urban micro-meteorological change can be simulated.
[0016] (3) Graph structure design
[0017] Each PV power station is regarded as a node in the graph, and the edges are defined according to the geographical locations and meteorological conditions between the power stations. A complete graph structure is constructed to describe the spatial dependency relationship between different PV power stations in the city.
[0018] (4) Time dynamic modeling
[0019] Long short-term memory network (LSTM) is used to analyze and process time series data. LSTM network is good at capturing long-term dependencies in time series and can effectively analyze the dynamic changes of meteorological data and photovoltaic power generation data in the time dimension.
[0020] Furthermore, for the graph convolutional neural network (GCN), the structure is as follows:
[0021] X′=GCN(X,A)
[0022] Among them, X is the feature matrix of meteorological data and power generation data, that is, the standardized meteorological data and power generation data, A is the adjacency matrix, which represents the spatial relationship between power stations, and X′ is the feature output of GCN.
[0023] The graph convolutional neural network (GCN) is used to process the standardized meteorological data and power generation data to obtain graph feature information A.
[0024] The calculation formula for node features in the l+1th layer of the graph convolutional neural network is as follows:
[0025]
[0026] in, is the adjacency matrix with self-loops added, I is the identity matrix, is the degree matrix, H (l) is the node feature of the lth layer, the node feature of the 0th layer is the preprocessed meteorological data and photovoltaic power generation data, W (l) is the weight matrix of this layer, σ is a nonlinear activation function (such as ReLU), and the total number of layers of the graph convolutional neural network can be defined artificially based on experience.
[0027] During the model training phase, the node features of the 0th layer are the pre-processed meteorological data and photovoltaic power generation data stored historically for modeling; during actual prediction, the node features of the 0th layer are the pre-processed meteorological data and photovoltaic power generation data stored recently.
[0028] During the model training phase, the initial value of the adjacency matrix A is defined as follows, where A ij =w ij :
[0029]
[0030] Among them, r is the threshold, d ij is the geographical distance between any two nodes i and j.
[0031] The adjacency matrix A represents the spatial relationship between power stations, which facilitates the spatial feature propagation and aggregation of the graph convolutional neural network (GCN). In the final model, the final value of the adjacency matrix A is obtained by model optimization.
[0032] Specifically, when designing the graph structure or optimizing the graph features, each PV power station is considered as a node in the graph. Each node contains its geographical location, historical power generation data, and meteorological data. Then, the geographical distance d between any two nodes i and j is calculated. ij , and decides whether to establish an edge between two nodes based on the preset distance threshold. The specific distance calculation method uses the spherical distance formula between two points on the earth's surface:
[0033] d ij = R arccos(sinφ i ·sinφ j +cosφ i ·cosφ j ·cos(λ i -λ j ))
[0034] Where R is the radius of the earth, φ i and φ j are the dimensions of nodes i and j respectively, λ i and λ j are the longitudes of nodes i and j respectively. The edge weight is attenuated according to the geographical distance and is represented by a Gaussian kernel function. ij If it is less than the threshold r, an edge is connected between i and j, and the weight of the edge is:
[0035]
[0036] When training the embedded graph convolution-long short-term memory network model, historical meteorological data and power generation data with time series data information are used as input, and the corresponding real photovoltaic power generation power is used as the supervision label for training.
[0037] Preferably, during modeling or actual prediction, the originally acquired meteorological data and power generation data need to be cleaned and standardized before being used as input for modeling or actual prediction.
[0038] Use GCN to get X′ corresponding to the data t, as the input feature of LSTM for final model training or data prediction. For each time step t, the input feature is X′ t LSTM network is used to capture the dynamic changes of time series data. The calculation formula of LSTM network is as follows:
[0039] H t =LSTM(X′ t , H t-1 )
[0040] Among them, X′ t is the input feature at time t, H t-1 is the hidden state of the previous moment, H t-1 is the hidden state at time t.
[0041] During the model training process, the standardized meteorological data and power generation data are used as input, the real power generation power P′ is used as the supervision label, and the root mean square error (RMSE) is used as the loss function to perform end-to-end training on the GC-LSTM model. As a preferred loss function, the RMSE is:
[0042]
[0043] Where n is the number of samples, P′ i is the actual power generation of the i-th sample, P′ i,pred is the predicted power generation of the i-th sample.
[0044] During the model training process, the back propagation algorithm is used to minimize the loss function and update the model parameters until the training converges.
[0045] Cascade GCN and LSTM to finally build a GC-LSTM hybrid model. The expression of the GC-LSTM hybrid model is as follows:
[0046] P′ pred =GCLSTM(X)
[0047] Among them, X is the standardized feature matrix of meteorological data and power generation data, P′ pred is the standardized power generation predicted by the model.
[0048] After the training converges, when the actual prediction is carried out, new meteorological data and power generation data are input, and the model is used to predict the power generation P′ of each power station in the future period of time. pred . The standardized predicted value P′ pred Denormalize to get the actual predicted value P pred The formula for denormalization is as follows:
[0049] P pred =P′pred ·P std +P mean
[0050] Among them, P std and P mean are the standard deviation and mean of the power generation data respectively.
[0051] The present invention combines GCN with LSTM to form an embedded hybrid model architecture, namely the GC-LSTM model. The model is trained end-to-end using standardized data as input. During the training process, the model will automatically learn the key meteorological factors affecting photovoltaic power generation and the interaction pattern of meteorological conditions between power stations. In the prediction stage, the model can infer the predicted power generation value of each power station in the future. Based on the prediction results, decision support can be provided for the load distribution and energy management of the power grid, energy utilization efficiency can be improved, refined scheduling can be achieved, and the safe, stable and efficient operation of the power system can be ensured.
[0052] The present invention combines the graph convolution-long short-term memory network (GC-LSTM) model to effectively improve the prediction accuracy of distributed photovoltaic power generation, and optimize the grid load distribution and energy management according to the prediction results. The present invention has the advantages of high prediction accuracy, strong anti-interference ability, high model lightweight, and fast computing speed, and can significantly improve the operating efficiency and reliability of distributed photovoltaic power generation systems under complex urban micro-meteorological conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 This is the technical roadmap of the distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model of the present invention. DETAILED DESCRIPTION
[0054] In order to more clearly explain the technical solution of the present invention, the prediction method and modeling process of the present invention will be described in detail below through specific embodiments. Specifically, the following steps are included:
[0055] (1) Data collection and preprocessing
[0056] First, all historical meteorological data and photovoltaic power generation data (i.e., power generation) collected from different power stations (e.g., n power stations: photovoltaic power station 1, photovoltaic power station 2, ... photovoltaic power station n) are cleaned to remove outliers and missing values. Then, the data is standardized to eliminate the impact of dimensions so that it meets the requirements of model training. The formula for standardization is as follows:
[0057]
[0058] Among them, x is the original data, μ is the mean of the data, and σ is the standard deviation of the data.
[0059] The historical database is established through step (1) for subsequent model training.
[0060] (2) Construction of urban micro-meteorological model
[0061] The urban micrometeorological model is mainly composed of a graph convolutional neural network (GCN) and a LSTM network.
[0062] The graph convolutional neural network (GCN) is used to process the standardized meteorological data and power generation data. The specific formula is as follows:
[0063] X′=GCN(X,A)
[0064] Among them, X is the feature matrix of meteorological data and power generation data, which are standardized in this embodiment, A is the adjacency matrix, which represents the spatial relationship between power stations, and X′ is the feature output by GCN. Using the graph convolutional neural network (GCN), the inherent spatial dependency characteristics of the input meteorological information (meteorological data and power generation data) and location information are constructed.
[0065] The calculation formula of the GCN layer is:
[0066] H (0) =X
[0067]
[0068] in, To add the adjacency matrix of self-loops, is the degree matrix, H (l) is the node feature of the lth layer, H (0) is the node feature of layer 0, H (1) is the node feature of the first layer, H (l+1) is the node feature of the l+1th layer, W (l) is the weight matrix to be learned in the (l) layer, and σ is a nonlinear activation function (such as ReLU).
[0069] Next, a detailed urban micro-meteorological model is constructed to ensure that the model can fully reflect the impact of the actual meteorological environment on photovoltaic power generation. The operation method is as follows:
[0070] Y=f model (X′)
[0071] Among them, f model is the model to be trained. Model f model The time series data will be further processed using LSTM networks to incorporate spatial and temporal dependencies to improve the accuracy of power forecasting.
[0072] (3) Design of graph structure
[0073] When designing the graph convolutional neural network (GCN), each photovoltaic power station is regarded as a node in the graph, and each node contains its geographical location, historical power generation data, meteorological data and other features. Then, the geographical distance d between any two nodes i and j is calculated. ij , and decides whether to establish an edge between two nodes based on the preset distance threshold. The specific distance calculation method uses the spherical distance formula between two points on the earth's surface:
[0074] d ij = R arccos(sinφ i ·sinφ j +cosφ i ·cosφ j ·cos(λ i -λ j ))
[0075] Where R is the radius of the earth, φ i and φ j are the dimensions of nodes i and j respectively, λ i and λ j are the longitudes of nodes i and j respectively. The edge weight is attenuated according to the geographical distance and is represented by a Gaussian kernel function. ij If it is less than the threshold r, an edge is connected between i and j, and the weight of the edge is:
[0076]
[0077] This weight definition makes the connection between nodes closer to each other stronger. Next, construct the adjacency matrix A of the graph, with an initial value A ij =w ij The adjacency matrix is defined as follows:
[0078]
[0079] The adjacency matrix A represents the spatial relationship between power stations, which facilitates the spatial feature propagation and aggregation of the graph convolutional neural network (GCN).
[0080] In summary, the graph feature X′ is obtained through the graph convolutional neural network (GCN) and used as input to construct the LSTM.
[0081] (4) Time dynamic modeling
[0082] The feature X′ output by GCN is used as the input feature of LSTM. For each time step t, the input feature is X′ t LSTM network is used to capture the dynamic changes of time series data. The calculation formula of LSTM network is as follows:
[0083] Ht =LSTM(X′ t , H t-1 )
[0084] Among them, X′ t is the input feature at time t, H t-1 is the hidden state of the previous moment. The following are the core components of the LSTM network and their functions:
[0085] ① Forget gate: determines the part of information that needs to be forgotten;
[0086] f t =σ(W f ·[h (t-1 ), x t ]+b f )
[0087] ② Input gate: determines the new information that needs to be input;
[0088] i t =σ(W i ·[h {t-1} , x t ]+b i )
[0089]
[0090] ③Memory unit update: update the memory unit at the current moment;
[0091]
[0092] ④Output gate: determines what information to output.
[0093] o t =σ(W o ·[H t-1 , X′ t ]+b o )
[0094] H t =o t tanh(C t )
[0095] Among them, f t ,i t , o t They are forget gate, input gate and output gate respectively, C t and H t are the unit state and hidden state, W and b are the corresponding gate network parameters.
[0096] (5) Model training and prediction
[0097] Cascade GCN and LSTM to build a GC-LSTM hybrid model. The expression of the GC-LSTM hybrid model is as follows:
[0098] P′ pred =GCLSTM(X)
[0099] Among them, X is the standardized feature matrix of meteorological data and power generation data, P′ pred is the standardized power generation predicted by the model.
[0100] The GC-LSTM model is trained end-to-end using the standardized meteorological data and power generation data as input, the real power generation P′ as the supervision label, and the root mean square error (RMSE) as the loss function. The specific training process includes the following steps:
[0101] ①Define the loss function RMSE:
[0102]
[0103] Where n is the number of samples, P′ i is the actual power generation of the i-th sample, P′ i,pred is the predicted power generation of the i-th sample.
[0104] ② Use the back-propagation algorithm and gradient descent optimization algorithm to minimize the loss function and update the model parameters until the training converges.
[0105] After the training converges, new meteorological data is input and the model is used to predict the power generation P′ of each power station in the future. pred . The standardized predicted value P′ pred Denormalize to get the actual predicted value P pred The formula for denormalization is as follows:
[0106] P pred =P′ pred ·P std +P mean
[0107] Among them, P std and P mean are the standard deviation and mean of the power generation data respectively.
[0108] Through the above method, the present invention can not only improve the accuracy of distributed photovoltaic power generation prediction, but also provide scientific data support for power grid operation, optimize power grid load management, and improve the overall operation efficiency and reliability of the system, and has broad application prospects.
Claims
1. A distributed photovoltaic power generation power prediction method based on an urban micro-meteorological model, characterized in that: include: By using the constructed embedded graph convolution-long short-term memory network, the recently acquired meteorological data and power generation data are used as input to realize the prediction of power generation power; the embedded graph convolution-long short-term memory network is composed of a graph convolutional neural network and a long short-term memory network, and the output of the graph convolutional neural network is used as the input of the long short-term memory network to realize the prediction of the distributed photovoltaic power generation power.
2. The distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model according to claim 1 is characterized in that: In the embedded graph convolution-long short-term memory network, the graph convolution neural network constructs feature graph information containing geographic location, meteorological data and power generation data; the long short-term memory network uses the feature graph information and the time series data dependency to obtain the predicted value of the power generation.
3. The distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model according to claim 1 or 2 is characterized in that: When training the embedded graph convolution-long short-term memory network model, historical meteorological data and power generation data with time series data information are used as input, and the corresponding real photovoltaic power generation power is used as the supervision label for training.
4. The distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model according to claim 1 or 2 is characterized in that: When modeling or actually predicting, the original meteorological data and power generation data need to be cleaned and standardized before being used as input for modeling or actually predicting.
5. The distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model according to claim 1 or 2 is characterized in that: The calculation formula for node features in the l+1th layer of the graph convolutional neural network is as follows: in, To add the adjacency matrix of self-loops, is the degree matrix, H (l) is the node feature of the lth layer, the node feature of the 0th layer is the preprocessed meteorological data and photovoltaic power generation data, W (l) is the weight matrix of this layer, and σ is the nonlinear activation function.
6. The distributed photovoltaic power generation power prediction method based on urban micro-meteorological model according to claim 4 is characterized in that: During the model training phase, the initial value of the adjacency matrix A is defined as follows, where A ij =w ij : Among them, r is the threshold, d ij is the geographical distance between any two nodes i and j.
7. The distributed photovoltaic power generation power prediction method based on the urban micro-meteorological model according to claim 2 is characterized in that: The embedded graph convolution-long short-term memory network model is trained end-to-end using root mean square error as the loss function.
8. The distributed photovoltaic power generation power prediction method based on urban micro-meteorological model according to claim 6 is characterized in that: The loss function RMSE is: Where n is the number of samples, P′ i is the actual power generation of the i-th sample, P′ i,pred is the predicted power generation of the i-th sample.
9. The distributed photovoltaic power generation power prediction method based on urban micro-meteorological model according to claim 2 is characterized in that: The back-propagation algorithm is used to minimize the loss function and update the model parameters until the training converges.
Citation Information
Cited By
Regional photovoltaic power direct interval prediction method based on graph convolutional neural network
CN119886427A
Photovoltaic module performance degradation prediction method and system based on GNN and LSTM
CN121211931A
Photovoltaic module performance degradation prediction method and system based on GNN and LSTM
CN121211931B
