A regional power generation prediction method and system based on dynamic random graph neural network

Abstract

The present invention relates to a regional power generation prediction method and system based on a dynamic random graph neural network. The method comprises the following steps: S1: collecting historical measured meteorological data and the power generation at the corresponding time for each power generation site in the region to construct historical measured data X'; S2: performing dynamic random modeling on each power generation site in the region to construct a dynamic representation of the power generation site, and constructing a dynamic random graph A based on the dynamic representation; S3: inputting X' and A into the dynamic random graph neural network, first passing them through a dynamic graph convolution layer to obtain a hidden state H representing the spatial correlation between the power generation sites; H passing through a temporal convolution layer to obtain a hidden state Z representing the temporal correlation; Z passing through two fully connected layers to generate future power prediction values ​​for each power generation site; and S4: constructing a loss function to train the dynamic random graph neural network. The method provided by the present invention can effectively learn the spatiotemporal correlation between power generation sites and improve the accuracy of power generation prediction.

Description

Technical Field

[0001] The present invention relates to the fields of new energy power generation and data mining, and in particular to a regional power generation prediction method and system based on a dynamic random graph neural network. Background Art

[0002] Amidst global energy resource shortages and increasingly prominent environmental issues, the research, development, and promotion of new energy technologies have become increasingly important. New energy technologies include clean and renewable energy sources such as solar, wind, hydro, and geothermal energy. Compared to traditional fossil fuels, these technologies are more environmentally friendly and sustainable. With the development of new energy technologies, their share of total grid-connected capacity has continued to increase. In China, new energy has become a crucial component of the national energy strategy. By the end of 2021, China's installed capacity of new energy had exceeded 900 million kilowatts, accounting for over 30% of the total. Among new energy technologies, the application of wind, photovoltaic, and hydropower technologies is becoming increasingly mature, and their power generation efficiency is gradually improving, effectively meeting people's demand for electricity. At the same time, the continuous advancement of new energy technologies has created new opportunities for technological innovation and industrial development. Countries are increasing their investment in the research and development of new energy technologies, promoting technological innovation and upgrading, and further boosting the development of the new energy industry. The Chinese government has also increased its support for new energy technologies, encouraging enterprises to strengthen technological research and development and market expansion, thereby enhancing the core competitiveness of the new energy industry.

[0003] Therefore, the research, development, and promotion of new energy technologies are of great significance. By developing new energy technologies, we can improve energy efficiency, reduce environmental pollution, protect the ecological environment, and promote sustainable development. In the future, the development of new energy technologies will become a key trend in global energy transformation and sustainable development, and will also bring new opportunities and challenges to the economic and social development of various countries.

[0004] However, regional renewable energy generation forecasting still faces numerous challenges. In southern my country, most wind farms are located in mountainous areas with complex terrain and climatic conditions. These locations, with their significant differences in altitude, often experience significant variations in power output. This leads to significant regional power output instability, making forecasting more challenging than for other energy sources. Photovoltaic power stations, on the other hand, experience seasonal fluctuations in power output and are heavily dependent on weather conditions. During cloudy seasons, these stations inevitably experience significant fluctuations. This volatility also presents challenges for grid scheduling, power balancing, and distribution. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a regional power generation prediction method and system based on dynamic random graph neural network.

[0006] The technical solution of the present invention is: a regional power generation prediction method based on a dynamic random graph neural network, comprising:

[0007] Step S1: Collect the historical measured weather data and the power generation at the corresponding time of each power generation site in the region, perform preprocessing, and construct the historical measured data X';

[0008] Step S2: Perform dynamic random modeling on each power generation site in the region, construct a dynamic representation of the power generation site, and construct a dynamic random graph A between the power generation sites based on the dynamic representation;

[0009] Step S3: Input X' and A into the dynamic random graph neural network. First, pass them through the dynamic graph convolution layer to obtain the hidden state H that represents the spatial correlation between power generation sites. H passes through the temporal convolution layer to obtain the hidden state Z that represents the temporal correlation. Z passes through two fully connected layers to generate the future power prediction value of each power generation site.

[0010] Step S4: constructing a loss function to train the dynamic random graph neural network.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] The present invention discloses a regional power generation prediction method based on a dynamic random graph neural network, and proposes a new dynamic random graph neural network for regional renewable energy power generation prediction. It can effectively learn the spatiotemporal correlation between power generation sites, and model the randomness existing in the observation data. It can also realize the adaptive fusion of the data of each power generation site based on the learned correlation between the power generation sites, thereby improving the accuracy of power generation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a method for regional power generation prediction based on a dynamic random graph neural network in an embodiment of the present invention;

[0014] Figure 2 Schematic diagram of the architecture of a dynamic random graph neural network in an embodiment of the present invention;

[0015] Figure 3 This is a structural block diagram of a regional power generation prediction system based on a dynamic random graph neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The present invention provides a regional power generation prediction method based on a dynamic random graph neural network, which can effectively learn the temporal and spatial correlations between power generation sites and improve the accuracy of power generation prediction.

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below through specific implementation and in conjunction with the accompanying drawings.

[0018] Example 1

[0019] like Figure 1 As shown, an embodiment of the present invention provides a regional power generation prediction method based on a dynamic random graph neural network, comprising the following steps:

[0020] Step S1: Collect the historical measured weather data and the power generation at the corresponding time of each power generation site in the region, perform preprocessing, and construct the historical measured data X';

[0021] Step S2: Perform dynamic random modeling on each power generation site in the region, construct a dynamic representation of the power generation site, and construct a dynamic random graph A between the power generation sites based on the dynamic representation;

[0022] Step S3: Input X' and A into the dynamic random graph neural network. First, pass them through the dynamic graph convolution layer to obtain the hidden state H that represents the spatial correlation between power generation sites. H passes through the temporal convolution layer to obtain the hidden state Z that represents the temporal correlation. Z passes through two fully connected layers to generate the future power prediction value of each power generation site.

[0023] Step S4: Construct a loss function to train the dynamic random graph neural network.

[0024] In one embodiment, the above step S1: collecting historical measured weather data and power generation at corresponding times at each power generation site in the region, performing preprocessing, and constructing historical measured data X' specifically includes:

[0025] Step S11: Collect the measured weather data and the power generation at the corresponding time at each power generation site in the region within a day. Each power generation site contains 6 features: 5 measured weather data and power generation, expressed as T d is the number of sampling points in a day; the measured meteorological data include: irradiance, temperature, humidity, air pressure, and wind speed;

[0026] Step S12: Collecting measured data of the power generation site for the historical period T Where T is the historical period; N is the number of power generation sites;

[0027] In the embodiment of the present invention, the sampling period in one day is set to 15 minutes, and there are 96 sampling points in one day, namely, T d =96. Collect the historical measured data of all power generation sites in the region within one week, that is, T = 7, and then we can get X = {x1, x2, ..., x i}∈RN×(7*96)×6 .

[0028] Step S13: Use the normalization method to scale the numerical value of X, expressed as X'=(X-μ x ) / σ x , where μ x ,σ x ∈R N×6 are the mean and variance vectors of X respectively.

[0029] In one embodiment, the above step S2: performing dynamic random modeling on each power generation site in the region, constructing a dynamic representation of the power generation site, and constructing a dynamic random graph A between the power generation sites based on the dynamic representation, specifically includes:

[0030] Step 21: For each time point t, define a mean embedding matrix and a covariance embedding matrix d is the hidden state dimension of the power generation site;

[0031] Each power generation site i is represented as a d-dimensional Gaussian distribution in, is the mean vector of power generation site i, which represents the position of power generation site i in the embedding space; is the covariance vector of power generation site i, which is used to characterize the uncertainty of power generation site i;

[0032] The dynamic modeling form of the power generation site is as follows:

[0033]

[0034]

[0035] Among them, T d is the number of sampling points in a day;

[0036] In actual scenarios, the state of renewable energy power generation sites is affected by many factors, such as geographical location, weather fluctuations, etc. These factors cause uncertainty. Through research, it is found that the state of power generation sites at the same time usually obeys a certain probability distribution, and this distribution changes over time. However, the existing methods cannot capture this uncertainty and dynamic characteristics by statically modeling each power generation site. Therefore, this paper models each power generation site as a dynamic random distribution. To integrate this uncertainty and dynamic characteristics.

[0037] Step S22: Calculate the two power generation sites and The Wasserstein distance between:

[0038]

[0039]

[0040]

[0041]

[0042]

[0043] Among them, W μ and W ∑ is a mapping matrix, ELU is an exponential linear unit, which maps the covariance vector to [-1, +∞); Trace() is a function for finding the trace of the matrix; ||||2 represents the L2 norm;

[0044] Existing graph neural network (GNN)-based methods typically use dot products to describe dependencies between power generation sites, but they are unable to measure differences between distributions. Furthermore, existing metrics often do not satisfy triangle inequalities and therefore cannot guarantee transitivity in the metric space. Transitivity, a key property in power generation forecasting, is crucial for effective modeling of power generation sites. To address this issue, this paper introduces the Wasserstein distance to measure the distance between random distributions. Compared to the dot product, the Wasserstein distance effectively incorporates uncertainty into spatiotemporal correlations while maintaining transitivity of similarity. Considering that the correlation between power generation sites is determined by current meteorological conditions and their location information, this paper combines the random embedding representation of the power generation sites with current meteorological conditions to measure the correlation between different power generation sites. By constructing a dynamic random graph between power generation sites, it is possible to effectively integrate independent meteorological data and observed historical power output data from different power generation sites, improving the accuracy of power generation forecasts. A smaller Wasserstein distance indicates more similar distributions between two power generation sites, which also indicates a stronger correlation between them. Therefore, the Wasserstein distance is used to define the adjacency matrix of regional power generation sites.

[0045] Step S23: Select the first k power generation sites with the closest Wasserstein distance to power generation site i as neighbor nodes of i, and construct the dynamic adjacency matrix A of the power generation sites:

[0046]

[0047]

[0048] Wherein, argtopk(·) is the index of the first k maximum values ​​in the vector, i.e., the power generation sites with the first k minimum Wasserstein distances to power generation site i; : represents all power generation sites except power generation site i;

[0049] A t,ij Represents the correlation between power generation site i and power generation site j at time t.

[0050] In this step, the first k power stations with the closest Wasserstein distance to power station i are selected as neighbors of i. For other power stations, their weights are set to zero to make the adjacency matrix sparse.

[0051] In one embodiment, in step S3 above, X' and A are input into a dynamic random graph neural network, first passing through a dynamic graph convolution layer to obtain a hidden state H representing the spatial correlation between power generation sites; H passes through a temporal convolution layer to obtain a hidden state Z representing the temporal correlation; Z passes through two fully connected layers to generate the future power prediction value of each power generation site. Specifically include:

[0052] Step S31: X' and the dynamic adjacency matrix A corresponding to time t t The input is a dynamic random graph neural network, which first undergoes a dynamic convolution operation in the dynamic graph convolution layer to obtain the hidden state H of the power generation site to represent the spatial correlation between the power generation sites:

[0053]

[0054]

[0055] Among them, H represents the hidden state of the initial state X' of the power station after passing through the dynamic random graph neural network; D t is the diagonal matrix, W k is the learnable parameter matrix;

[0056] As a representative approach, traditional graph convolutional networks (GCNNs) demonstrate significant capabilities in handling spatial correlations. By aggregating neighbor information based on the topological structure of power plants, they can learn the spatial correlations between them. However, these correlations are dynamic, influenced by real-time meteorological conditions. Static graphs cannot capture their time-varying nature. To address this limitation, the present invention incorporates these time-varying correlations into the spatial correlations between power plants by constructing a dynamic graph convolutional layer.

[0057] Step S32: Input the hidden state H into the temporal convolution layer, and after the dilated convolution operation, obtain the hidden state Z to represent the temporal correlation of the power generation sites;

[0058] Z=tanh(Θ f *H)⊙σ(Θ g *H)

[0059] The symbol ⊙ represents the Hadamard product of element-by-element multiplication, tanh(·) and σ(·) are the hyperbolic tangent function and sigmoid function, respectively, used as filter and gating functions; * represents the dilated convolution operation, Θ f and Θ g is the corresponding convolution kernel of the convolution filter;

[0060] Although the dynamic graph convolution in step S31 has effectively modeled the spatial dependencies between power station locations, the long-term sequence trends of a location are also important. The future output power of a power station is highly correlated with its historical data. To learn the temporal correlation in time series, this paper constructs a temporal convolution layer and uses dilated convolution to process in the time domain, thereby capturing the temporal correlation of power stations.

[0061] Step S33: Pass Z through two fully connected layers as the prediction layer to generate the power prediction value of each power generation site at the future time

[0062]

[0063] Among them, W 1 and W 2 is the weight matrix of the fully connected network, b 1 and b 2 is the offset matrix of the fully connected network, q is the number of steps in the future, It represents the predicted power generation of N power generation sites at a total of q time points from the future time t+1 to the future time t+q.

[0064] In one embodiment, the above step S4: constructing a loss function to train a dynamic random graph neural network specifically includes:

[0065] Constructing a loss function based on MAE To train a dynamic random graph neural network:

[0066]

[0067] Among them, Y t+i,j is the actual value of the power generation of power station j at the t+i moment, is the predicted power generation value of power generation site j at time t+i.

[0068] Figure 2 Shows a schematic diagram of the architecture of a dynamic random graph neural network.

[0069] The present invention discloses a regional power generation prediction method based on a dynamic random graph neural network, and proposes a new dynamic random graph neural network for regional renewable energy power generation prediction. It can effectively learn the spatiotemporal correlation between power generation sites, and model the randomness existing in the observation data. It can also realize the adaptive fusion of the data of each power generation site based on the learned correlation between the power generation sites, thereby improving the accuracy of power generation prediction.

[0070] Example 2

[0071] like Figure 3 As shown, an embodiment of the present invention provides a regional power generation prediction system based on a dynamic random graph neural network, including the following modules:

[0072] The data preprocessing module 51 is used to collect the historical measured weather data and the power generation at the corresponding time of each power generation site in the region, perform preprocessing, and construct the historical measured data X';

[0073] A dynamic random graph construction module 52 is used to perform dynamic random modeling on each power generation site in the region, construct a dynamic representation of the power generation site, and construct a dynamic random graph A between the power generation sites based on the dynamic representation;

[0074] The power generation prediction module 53 inputs X' and A into the dynamic random graph neural network. First, it passes through the dynamic graph convolution layer to obtain the hidden state H that represents the spatial correlation between power generation sites. H passes through the time convolution layer to obtain the hidden state Z that represents the temporal correlation. Z passes through two layers of fully connected layers to generate the future power prediction value of each power generation site.

[0075] The loss function construction module 54 is used to construct a loss function to train the dynamic random graph neural network.

[0076] The above embodiments are provided for the purpose of describing the present invention only and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present invention are intended to be within the scope of the present invention.

Claims

1. A regional power generation prediction method based on dynamic random graph neural network, characterized in that: include: Step S1: Collect the historical measured weather data of each power generation site in the region and the power generation at the corresponding time, perform preprocessing, and construct the historical measured data. ; Step S2: Perform dynamic random modeling on each power generation site in the region, construct a dynamic representation of the power generation site, and construct a dynamic random graph between the power generation sites based on the dynamic representation. , specifically including: Step 21: For each time point , define a mean embedding matrix and a covariance embedding matrix , is the hidden state dimension of the power generation site; Each power generation site Characterized as a dimensional Gaussian distribution ,in, For power generation sites The mean vector of Position in the embedding space; For power generation sites The covariance vector of uncertainty; The dynamic modeling form of the power generation site is as follows: in, is the number of sampling points in a day; Step S22: Calculate the two power generation sites and The Wasserstein distance between: in, and is the mapping matrix, is an exponential linear unit that maps the covariance vector to ; is the function for finding the trace of a matrix; || ||2 represents the L2 norm; Step S23: Select a power generation site The Wasserstein distance to the nearest Power generation sites are used as neighbor nodes of 𝑖 to build a dynamic adjacency matrix of power generation sites : in, Vector Center Front The index of the maximum value, that is, the index of the Wasserstein distance to the power station i The power generation site with the minimum value; : represents all power generation sites except power generation site i; represents the correlation between power generation site i and power generation site j at time t; Step S3: and The input dynamic random graph neural network first passes through the dynamic graph convolution layer to obtain the hidden state that represents the spatial correlation between power generation sites. ; After the time convolution layer, the hidden state representing the temporal correlation is obtained ;Will After two fully connected layers, the future power forecast values ​​of each power generation site are generated ; Step S4: constructing a loss function to train the dynamic random graph neural network.

2. The regional power generation prediction method based on dynamic random graph neural network according to claim 1 is characterized in that: Step S1: Collect the historical measured weather data of each power generation site in the region and the power generation at the corresponding time, perform preprocessing, and construct the historical measured data. , specifically including: Step S11: Collect the measured weather data and the power generation at the corresponding time at each power generation site in the region within a day. Each power generation site contains 6 features: 5 measured weather data and power generation, expressed as , is the number of sampling points in a day; wherein the measured meteorological data include: irradiance, temperature, humidity, air pressure, and wind speed; Step S12: Collecting measured data of the power generation site for the historical period T ,in, For historical cycles; is the number of power generation sites; Step S13: Use normalization method to Scale the numerical value, expressed as ,in, They are The mean and variance vector of .

3. The regional power generation prediction method based on dynamic random graph neural network according to claim 2 is characterized in that: Step S3: and The input dynamic random graph neural network first passes through the dynamic graph convolution layer to obtain the hidden state that represents the spatial correlation between power generation sites. ; After the time convolution layer, the hidden state representing the temporal correlation is obtained ;Will After two fully connected layers, the future power forecast values ​​of each power generation site are generated , specifically including: Step S31: and the dynamic adjacency matrix corresponding to time t Input the dynamic random graph neural network, first pass the dynamic convolution operation of the dynamic graph convolution layer to obtain the hidden state of the power station , to characterize the spatial correlation between power generation sites: in, Represents the initial state of the power generation site The hidden state after passing through the dynamic random graph neural network; is the diagonal matrix, is the learnable parameter matrix; Step S32: The hidden state Input the time convolution layer, after the expansion convolution operation, the hidden state is obtained , to characterize the temporal correlation of power generation sites; Among them, the symbol represents the Hadamard product of element-by-element multiplication, and They are the hyperbolic tangent function and the sigmoid function, used as filter and gating functions respectively; represents the dilated convolution operation, and is the corresponding convolution kernel of the convolution filter; Step S33: After two fully connected layers as prediction layers, the power forecast values ​​of each power generation site at future times are generated : in, and is the weight matrix of the fully connected network, and is the offset matrix of the fully connected network, The number of steps in the future. Indicates the time from the future time t+1 to the future time t+q, a total of q moments The predicted power generation capacity of each power generation site.

4. The regional power generation prediction method based on dynamic random graph neural network according to claim 3 is characterized in that: The step S4: constructing a loss function to train the dynamic random graph neural network specifically includes: Constructing a loss function based on MAE , to train the dynamic random graph neural network: in, is the actual value of power generation at power station j at time t+i, is the predicted power generation value of power generation site j at time t+i.

5. A regional power generation prediction system based on dynamic random graph neural network, characterized in that: Includes the following modules: The data preprocessing module is used to collect the historical measured meteorological data of each power generation site in the region and the power generation at the corresponding time, perform preprocessing, and construct historical measured data. ; Construct a dynamic random graph module to perform dynamic random modeling on each power generation site in the region, construct a dynamic representation of the power generation site, and construct a dynamic random graph between the power generation sites based on the dynamic representation , specifically including: Step 21: For each time point , define a mean embedding matrix and a covariance embedding matrix , is the hidden state dimension of the power generation site; Each power generation site Characterized as a dimensional Gaussian distribution ,in, For power generation sites The mean vector of Position in the embedding space; For power generation sites The covariance vector of uncertainty; The dynamic modeling form of the power generation site is as follows: in, is the number of sampling points in a day; Step S22: Calculate the two power generation sites and The Wasserstein distance between: in, and is the mapping matrix, is an exponential linear unit that maps the covariance vector to ; is the function for finding the trace of a matrix; || ||2 represents the L2 norm; Step S23: Select a power generation site The Wasserstein distance to the nearest Power generation sites as Neighbor nodes of the power generation site are constructed to form a dynamic adjacency matrix : in, Vector Center Front The index of the maximum value, that is, the index of the Wasserstein distance to the power station i The power generation site with the minimum value; : represents all power generation sites except power generation site i; represents the correlation between power generation site i and power generation site j at time t; Predicting the power generation module, and The input dynamic random graph neural network first passes through the dynamic graph convolution layer to obtain the hidden state that represents the spatial correlation between power generation sites. ; After the time convolution layer, the hidden state representing the temporal correlation is obtained ;Will After two fully connected layers, the future power forecast values ​​of each power generation site are generated ; A loss function module is constructed to construct a loss function to train the dynamic random graph neural network.

Images