Method, system and equipment for predicting aggregation adjustable capability of virtual power plant based on cloud-edge collaboration mechanism

Abstract

The invention discloses a method, a system and equipment for predicting the aggregation adjustable capability of a virtual power plant based on a cloud-side cooperation mechanism, and the method comprises the steps: carrying out the correlation analysis and feature reconstruction of input data based on a cloud-side cooperation system network system according to a maximum information coefficient, and carrying out the pre-preprocessing of the data; point prediction is carried out on a future aggregation adjustable capability value through a CNN-BiLSTM-QR model, probability intervals of the adjustable capability under different quantiles are reliably represented through a probability prediction model of quantile regression, and a point prediction result and an interval prediction result of the aggregation adjustable capability are output at the same time. According to the method, the adjustable capability of the demand side resources in the virtual power plant can be accurately predicted, so that the virtual power plant effectively participates in power system scheduling, and the operation efficiency and stability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the field of virtual power plant energy, and in particular to a method, system and device for predicting the aggregated adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism. Background Art

[0002] Accurate prediction of the adjustable capacity of demand-side resources in virtual power plants is an important part of effectively participating in power system dispatch and improving system operation efficiency and stability. The prediction of the adjustable capacity of virtual power plants is based on historical data of influencing factors such as user electricity consumption behavior, climate conditions, region, and equipment status, and studies the intrinsic connection between adjustable capacity and related factors. According to the relevant laws found in historical data, the trend of future adjustable capacity is predicted. However, there are two main methods for predicting the adjustable capacity of virtual power plants. One is load prediction considering demand response, which is an indirect prediction, and the other is direct prediction. Both of the above prediction methods are single-value predictions based on historical data, and only predict the response power index of adjustable loads. The uncertainty of influencing factors such as baseline load, weather environment, user response equipment, and the uncertainty of user adjustable capacity are not considered. At the same time, the prediction of the adjustable capacity of virtual power plants is mostly point prediction, which cannot fully reflect the uncertainty of data and will inevitably produce errors in the prediction results. Summary of the invention

[0003] Purpose of the invention: The present invention aims to provide a method for predicting the adjustable capacity of a virtual power plant that can improve accuracy; another purpose of the present invention is to provide a prediction system and device for the aggregated adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism.

[0004] Technical solution: The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism of the present invention comprises the following steps:

[0005] (1) Build a cloud-edge collaborative network system for regulating virtual power plants;

[0006] (2) Preprocess the collected edge node data, select features for each edge node using the maximum information coefficient method, and obtain a data set as the input of the corresponding edge node in the cloud-edge collaborative network system. The edge node performs regional adjustable capacity evaluation;

[0007] (3) Construct a CNN-BiLSTM-QR model to input the input data set of each edge node obtained from the cloud and the output regional adjustable capacity evaluation value to obtain the predicted value of the aggregated adjustable capacity of the virtual power plant; wherein the CNN-BiLSTM-QR model includes a local CNN layer, a local pooling layer, a global CNN layer, a global pooling layer, a BiLSTM layer, a fully connected layer and an output layer; the local CNN layer is used to input the input data set of the edge node and the output regional adjustable capacity evaluation value to obtain the data features of each edge node and the region described therein; the global CNN layer is used to input the data features of the edge nodes and the regions described therein output by all local pooling layers to obtain the global data features of the cloud layer; the loss function of the BILSTM neural network adopts a quantile loss function, sets multiple quantiles and predicts the adjustable capacity at each quantile, and then maps the data features to an output layer consistent with the number of quantiles through a fully connected layer as the output result, and selects the output at a specific quantile whose quantile loss function value is equal to the original loss function as the point prediction result output, and constructs prediction intervals of different confidence levels in the quantile sequence, and outputs the output result of the interval as a probability interval prediction result.

[0008] Furthermore, the edge node data collected include user baseline load, number of each type of users in the node, maximum allowed delay, electricity price, date, real-time temperature and daily maximum temperature.

[0009] Furthermore, the BiLSTM layer in step (3) introduces an attention mechanism. First, the BiLSTM hidden state sequence h = {h1,h2,…,h t} and x={x1,x2,…,x t}, attention weight α ij for

[0010]

[0011] e ij =tanh(ω h h i +ω s x j +b c )

[0012]

[0013] Among them, e ij and e ik represents the hidden state h of BiLSTM i And the global pooling layer output x j The attention score between j Represents the hidden state of the attention-weighted BiLSTM, which is used to represent the important information of the input sequence.

[0014] Furthermore, the quantile loss function QLOSS in step (3) is

[0015]

[0016] Where τ represents the quantile, Y i represents the actual adjustable capacity of the virtual power plant, f(·) represents the solution function of the estimated value of the response variable at the quantile τ under the condition of given characteristic explanatory variables, X represents the given characteristic explanatory variables, W(t) and V(t) represent the weight matrices between the hidden layer and the input and output layers of the neural network, respectively, ω i and v i Respectively represent the i-th element in the W(t) and V(t) matrices, and l represents the penalty parameter;

[0017] Set multiple quantiles and predict the adjustable capacity at different quantiles separately, and map various data features to the output layer consistent with the number of quantiles through the fully connected layer as the output result. Select the output at a specific quantile whose quantile loss function value is equal to the original loss function as the point prediction result output. Construct prediction intervals with different confidence levels by boxing in the quantile sequence, and output the output result of the interval as the probability interval prediction result;

[0018] The result output Q of the CNN-BiLSTM-QR model τ for

[0019] Q τ (Y|X)=f(X,W(τ),V(τ)).

[0020] Furthermore, the Monte Carlo sampling method is used to generate samples of M prediction scenarios, which are input into the trained CNN-BiLSTM-QR model to obtain the predicted value of the aggregated adjustable capacity of the virtual power plant. The predicted value of the aggregated adjustable capacity of the virtual power plant is corrected through the kernel density estimation probability density function to obtain the final predicted value of the aggregated adjustable capacity of the virtual power plant.

[0021] Furthermore, the probability density function for

[0022]

[0023] in, Indicates that at the quantile τ i where γ is the number of quantiles, h is the window width, and K(·) is the Gaussian kernel function.

[0024] Furthermore, the method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism also includes:

[0025] (5) The performance of the final prediction results is evaluated through point prediction evaluation indicators and confidence interval evaluation indicators. The evaluation indicators of point prediction include mean absolute error, root mean square error and mean absolute percentage error.

[0026] The prediction system of the aggregate adjustable capacity of a virtual power plant based on the cloud-edge collaboration mechanism of the present invention includes:

[0027] System building module, used to build a cloud-edge collaborative network system for regulating virtual power plants;

[0028] The data collection and processing module is used to collect edge node data and perform preprocessing. It selects features for each edge node through the maximum information coefficient method to obtain a data set as the input of the corresponding edge node in the cloud-edge collaborative network system. The edge node performs regional adjustable capacity evaluation;

[0029] The model building module is used to build a CNN-BiLSTM-QR model, which is used to input the input data set of each edge node obtained from the cloud and the output regional adjustable capacity evaluation value to obtain the predicted value of the aggregate adjustable capacity of the virtual power plant. The CNN-BiLSTM-QR model includes a local CNN layer, a local pooling layer, a global CNN layer, a global pooling layer, a BiLSTM layer, a fully connected layer and an output layer; the local CNN layer is used to input the input data set of the edge node and the output regional adjustable capacity evaluation value to obtain the data features of each edge node and the area described by it; the global CNN layer is used to input all local The data features of the edge nodes and the areas described in the partial pooling layer output are used to obtain the global data features of the cloud layer; the loss function of the BILSTM neural network adopts the quantile loss function, sets multiple quantiles, predicts the adjustable capabilities at different quantiles separately, and maps various data features to the output layer consistent with the number of quantiles through the fully connected layer as the output result, selects the output at a specific quantile whose quantile loss function value is equal to the original loss function as the point prediction result output, and constructs prediction intervals with different confidence levels by boxing in the quantile sequence, and outputs the output result of the interval as the probability interval prediction result.

[0030] The computer device of the present invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the steps of the above method are implemented when the processor executes the computer program.

[0031] The computer-readable storage medium of the present invention stores a computer program thereon, and the computer program implements the steps of the above method when executed by a processor.

[0032] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. The present invention performs a correlation analysis on the input data containing delay characteristics in the probability prediction of the aggregated adjustable capacity of the virtual power plant according to the maximum information coefficient, and effectively increases the accuracy of the prediction model after feature selection through the maximum information coefficient; 2. The present invention combines the CNN-BiLSTM-QR model with the cloud-edge collaboration mechanism and the attention mechanism to better capture the spatiotemporal connection of the internal data of each node, improves the ability to capture local features, and the point prediction algorithm proposed in the comparison algorithm has the highest prediction accuracy; 3. The present invention introduces quantile regression and kernel density estimation to obtain the point prediction results and probability interval prediction results of the adjustable capacity at different quantile points and the probability density curve. The probability prediction algorithm proposed in the comparison algorithm has the highest prediction accuracy and the fastest prediction speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention;

[0034] Figure 2 It is a structural diagram of the CNN-BiLSTM-QR model;

[0035] Figure 3 Heatmap for correlation analysis between regional nodes;

[0036] Figure 4 This is the prediction result diagram of the aggregated peak shaving capacity points;

[0037] Figure 5 This is the prediction result diagram of the aggregated valley filling capacity point;

[0038] Figure 6 This is the prediction result diagram of the aggregate peak-cutting capacity at each quantile point;

[0039] Figure 7 This is the probability prediction result diagram when the confidence level is 90%;

[0040] Figure 8 This is the prediction result diagram of the probability density curve at different times. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings.

[0042] The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism of the present invention comprises the following steps:

[0043] (1) Build a cloud-edge collaborative network system for regulating virtual power plants;

[0044] The architecture of the cloud-edge collaborative network system of the virtual power plant described in the present invention is different from the centralized architecture. The edge nodes are placed close to the user side. The central cloud relies on the virtual power plant cloud to centrally integrate the computing resources of many servers and is responsible for processing global high-complexity data services, including comprehensive evaluation and prediction of the adjustable aggregation capacity of the virtual power plant, decomposition of scheduling instructions to each edge node, and overall communication optimization control of the cloud-edge-end system. Technical service providers and load aggregators are located at the edge nodes. The edge servers they use have the characteristics of low energy consumption, low latency, and fast reading and writing speeds, and are suitable for computing terminal-oriented tasks with high reliability and low latency requirements. Data from each layer is collected and collected in the cloud center. By calculating and analyzing the collected data, the cloud layer handles the scheduling control and external cooperation services of the virtual power plant from the system level. During the scheduling process at all levels, the model center of the cloud layer is responsible for storing and calculating global, highly complex models that aggregate all edge layers, thereby ensuring global optimization scheduling within the virtual power plant. The model center of the edge layer is responsible for calculating the aggregated load resource model near the edge node to share the computing pressure of the cloud center, reduce the large amount of time delay caused by the direct interaction between the cloud layer and the terminal data in centralized scheduling, and ensure the data privacy security of virtual power plant service users.

[0045] Cloud layer: The virtual power plant cloud server has rich storage and computing resources and can interact with platforms such as the power dispatch center, power grid cloud platform, and power grid data center. In the comprehensive evaluation process of the aggregated adjustable capacity of the virtual power plant, the cloud server is mainly responsible for using the communication system optimization method to plan the information transmission path for the network data within the system to optimize the latency and reliability of data transmission in the information system, and evaluate the aggregated adjustable capacity of the virtual power plant that takes into account both communication and action time based on the uploaded aggregate terminal cluster data. In the process of adjustable capacity prediction, the edge evaluation data is transmitted to the edge prediction model, and the cloud evaluation data is input into the aggregated adjustable capacity prediction model at the same time, and the cloud-edge mechanism is used to realize refined adjustable capacity prediction. Based on external data such as electricity price data and unit data, the cloud platform performs calculations for the economic dispatch of the virtual power plant participating in the power system to ensure economic dispatch based on the balance of supply and demand. Then, based on the obtained day-ahead and intraday dispatch plans, the optimal dispatch of the adjustable resources within the virtual power plant is guided.

[0046] Edge layer: The edge layer plays the role of connecting the terminal and the cloud layer in cloud-edge collaboration. The edge node has both the uplink communication capability with the virtual power plant cloud platform and the downlink communication capability with the distribution substation, including edge controllers, edge servers, and edge gateways. Each edge node is not only responsible for the data processing and operation management of the distribution substation area where the regional data is uploaded, but also has the function of assisting the analysis of large models deployed on the cloud server in the evaluation and prediction of the adjustable capacity of edge aggregation. In the evaluation of the adjustable capacity of virtual power plant aggregation, the edge node is responsible for collecting the uploaded user operation data, evaluating the adjustable capacity of users in the region based on the differentiated evaluation model configured on the edge side, and uploading the cluster data, which is further evaluated by the cloud platform. In addition, edge nodes can also be configured with small user-oriented applications, such as short-term load baseline forecasting, demand response plan filling, etc., to help users fine-tune future electricity consumption behaviors.

[0047] Terminal: The response terminal equipment is deployed on the user side, responsible for real-time collection of power data and equipment operation data and uploading them to the distribution substation. The user terminal and the distribution substation are in a many-to-one correspondence. The terminal is mainly intelligent-oriented, controls multiple deployed intelligent systems, and comprehensively regulates the operation of various intelligent electrical equipment according to the downloaded terminal control instructions.

[0048] (2) The collected edge node data is preprocessed, and the features of each edge node are selected through the maximum information coefficient method to obtain a data set, which serves as the input of the corresponding edge node in the cloud-edge collaborative network system. The edge node performs regional adjustable capacity evaluation.

[0049] The edge node data is preprocessed as follows:

[0050] The reliability of input data has an important impact on the adjustable capacity prediction model. In order to ensure the accuracy of the adjustable capacity probability prediction, the model's original data needs to be preprocessed, including abnormal data detection, missing value filling and normalized data.

[0051] Abnormal data detection. Because abnormal data differ greatly from normal data, their existence may cause large errors in the model. Drawing a box plot can help identify abnormal values ​​in the data. The box plot shows the distribution of the data and identifies abnormal values. At the same time, it refers to the fluctuation range of normal data. If the data exceeds the specified range, it should be discarded and regarded as missing values.

[0052] Missing value filling. Since time series data is recorded for a long time, when equipment fails or power outages occur, missing data will be generated. The missing data will affect the coherence of the time series and may also cause the input data to be misplaced. When there is less missing data, the data at the previous moment is used as the data within that period. When the missing range is large, interpolation is used to fill in the data.

[0053] Normalize data. Since different data have different units, the data of various categories may vary greatly, which will lead to problems such as slow training and local convergence. It is necessary to normalize various types of data. The formula is as follows:

[0054]

[0055] In the formula, x' is the normalized variable, x max and x min are the maximum and minimum values ​​of the variable x in the sample, respectively. After using the model to predict the aggregate adjustable capacity, the predicted adjustable capacity value is denormalized.

[0056] The maximum information coefficient method is used to select features for each edge node, as follows:

[0057] The essence of the prediction task is to find the rules and changing trends between the input data and the output data, which requires that the input data and the output data have a correlation. The size of the correlation determines which data is most important for the evaluation of adjustable capacity. Correlation analysis can help determine which features need to be input and reduce unnecessary features. The present invention selects the user baseline load in each node area, the number of users of each type in the node, the maximum allowable delay, electricity price, time and temperature information as the input sample set, and performs correlation analysis on the input sample set. Most studies use Pearson correlation coefficient, Spearman correlation coefficient or Kendall correlation coefficient as a method for analyzing the correlation between data. However, Pearson correlation coefficient cannot be applied to nonlinear data, and Spearman and Kendall correlation coefficients are only applicable to nonlinear data with monotonicity. The maximum information coefficient (MIC) is a method that divides the data space into grids, calculates the correlation of variables in each grid, and selects the grid with the largest result value as the maximum information coefficient value. It is suitable for more complex nonlinear data and has the characteristics of high robustness. The maximum information coefficient M∈[-1,1], when the value is 0, it means that the variable X and the variable Y are independent of each other, and when the value is 1, it means that X and Y are completely correlated. The formula can be expressed as:

[0058]

[0059] Where I(X,Y) represents the mutual information between variables X and Y, which is used to measure the degree of mutual influence between X and Y. p(x,y) is the joint probability distribution function, p(x) and p(y) are marginal probability distribution functions, and S and T are subsets of X and Y, respectively.

[0060] When using the CNN-BiLSTM-QR network combined with the cloud-edge mechanism to predict adjustable capacity, it is first necessary to analyze the correlation between the input variables and the adjustable capacity and reconstruct the input features. The input features considered in the present invention are divided into two categories. One is the electrical characteristics that affect the aggregated adjustable capacity of the virtual power plant, including user baseline load, the number of users of each type in the node, the maximum allowable delay and electricity price; the other is non-electrical characteristics, including date, real-time temperature and daily maximum temperature. Most of these data are highly random, and the maximum information coefficient method performs well in capturing the nonlinear relationships therein. Since the prediction model of the present invention is combined with the cloud-edge collaborative information architecture, the user baseline load data input into the cloud-side prediction model is divided by the network into regional baseline load data of each edge node, and each regional node j contains a user baseline load sequence Temperature series For k-dimensional input data, the edge node uses the maximum information coefficient to evaluate the correlation between the current category data and the edge node aggregation adjustable capacity, and selects the features contained in the a regional nodes with the highest correlation with the current category data as the input of the edge node. Each edge node receives a total of ak feature sequences as input, and each regional node data can be selected by multiple edge nodes at the same time.

[0061] (3) Construct a CNN-BiLSTM-QR model, which is used to input the input data set of each edge node obtained from the cloud and the output regional adjustable capacity evaluation value to obtain the predicted value of the aggregated adjustable capacity of the virtual power plant; wherein the CNN-BiLSTM-QR model includes a local CNN layer, a local pooling layer, a global CNN layer, a global pooling layer, a BiLSTM layer, a fully connected layer and an output layer; the local CNN layer is used to input the input data set of the edge node and the output regional adjustable capacity evaluation value to obtain the data features of each edge node and the region described therein; the global CNN layer is used to input the data features of the edge nodes and the regions described therein output by all local pooling layers to obtain the global data features of the cloud layer; the loss function of the BILSTM neural network adopts a quantile loss function, sets multiple quantiles and predicts the adjustable capacity at different quantiles separately, and maps various data features to an output layer consistent with the number of quantiles through a fully connected layer as the output result, selects the output at a specific quantile whose quantile loss function value is equal to the original loss function as the point prediction result output, and constructs prediction intervals of different confidence levels by boxing in the quantile sequence, and outputs the output result of the interval as the probability interval prediction result output.

[0062] (4) The CNN network has good feature extraction capabilities and is used to capture the characteristic relationship between input data and predict adjustable capacity. The BiLSTM network is a neural network that trains the forward and backward propagation of adjustable capacity time series. It can discover the connection between the adjustable capacity at a certain point and past and future data. QR is used to reliably represent the probability interval of adjustable capacity at different quantiles. The results of the virtual power plant adjustable potential assessment are divided into two parts as cloud and edge input data. One part is the regional adjustable capacity results within the range of the regional node to which the edge node assesses, and the environmental data and load historical operation data are used as edge data input to the corresponding CNN edge model; the other part is a comprehensive cloud-based assessment of the aggregated adjustable capacity of the virtual power plant, including the adjustable capacity assessment results of each regional node under different maximum allowable delays, which are used as cloud data input to the CNN corresponding to each edge node in the CNN-BiLSTM-QR model.

[0063] The upper layer of the CNN-BiLSTM-QR model is the local CNN to which the edge node belongs, the middle layer is the global CNN and the BiLSTM model for extracting time series features, and the lower layer is the fully connected layer and the QR model for calculating the adjustable capacity at each quantile. After the cloud-based probability prediction model obtains the input data, the data of the corresponding edge node is sent to the convolution layer and pooling layer of the corresponding CNN to extract the spatiotemporal features of the data of each edge node and the regional node. The output data features of each edge node are extracted by the independent convolution layer and pooling layer of the cloud. The CNN used to extract global data flattens the output data of the corresponding CNN of each node and sends it to the BiLSTM network to extract the time series features in the data. The time series data passes through the forward propagation hidden layer and the backward propagation hidden layer respectively, and the hidden layer parameters are trained from the past to the future and from the future to the past. The two results are spliced ​​to obtain the output results of BiLSTM. Through the fully connected layer, the multi-dimensional mapping features are converted into one-dimensional probability prediction output results. The number of output layers depends on the number of quantiles set based on the QR mechanism. Finally, the aggregated adjustable capacity prediction values ​​at different quantiles are obtained by merging and merging into prediction intervals.

[0064] The CNN-BiLSTM-QR prediction framework combined with the cloud-edge mechanism of the present invention inputs the data of each edge node into a separate local CNN model, which is conducive to better capturing the spatial connection of the internal data of each node. Each local CNN can focus on identifying the characteristics of the specific user response behavior and spatial environment related to each node, thereby improving the learning ability of local features. The outputs of these local CNN models are connected to a shared CNN-BiLSTM model to capture the overall temporal characteristics and global data dynamic relationships between nodes, thereby improving the learning ability of global features. At the same time, the cloud-edge mechanism decomposes the large CNN model in the traditional prediction framework into multiple small models, which helps to reduce the overall number of parameters and speed up the model training time. The smaller parameter amount of the local CNN model corresponding to each edge node also helps to reduce the risk of overfitting. When some edge nodes change, the model combined with the cloud-edge mechanism has greater flexibility than the traditional model.

[0065] The CNN-BiLSTM point prediction model with the introduction of the attention mechanism is as follows:

[0066] The CNN-BiLSTM model is a deep learning model that combines convolutional neural networks and bidirectional long short-term memory networks, and is used to process sequence data with spatiotemporal characteristics. In the prediction of the adjustable capacity of virtual power plants, the model can effectively capture the spatiotemporal characteristics of input load-related data and predict future adjustable capacity through the trained model. In the CNN-BiLSTM model, CNN is used to extract spatial features from the input data, such as the historical adjustable capacity features of each node. The convolution layer and pooling layer in its structure are used for feature extraction and dimensionality reduction of load data in the input layer, and the fully connected layer is used to map the extracted features to the output data. BiLSTM learns the historical information and future trends of the data through forward and backward propagation, respectively, so as to better predict future adjustable capacity. Its LSTM unit structure consists of an input gate i t 、Forget gate t , unit state g t and the output part o t The calculation process is as follows:

[0067] i t =s(ω i [h t-1 ,p t ]+b i )

[0068] f t =s(ω f [h t-1 ,p t ]+b f )

[0069] g t=tanh(ω g [h t-1 ,p t ]+b f )

[0070] r t =f t ⊙r t-1 +i t ⊙g t

[0071] o t =s(ω o [h t-1 ,p t ]+b o )

[0072] h t =o t ⊙tanh(r t )

[0073] In the formula, ω and b are the parameters of the network model, representing weight and bias, and h t and r t Represents hidden state and updated state, x t is the input load processing data at time t, s is the sigmoid function used to map real numbers to probability values, and ⊙ represents the Hadamard product.

[0074] The attention mechanism is a technology that dynamically focuses on different parts of the input data sequence when processing sequence data. The basic idea is to dynamically assign different attention weights to different sequence elements based on the content of the input sequence data and the relationship between the previous and next data, so that the model can better understand and utilize the important information in the sequence data. In the adjustable capacity prediction, the attention mechanism allows the model to dynamically focus on different parts of the input related data that affect the adjustable capacity when processing sequence data. It can help the model focus on the load characteristics that have an important impact on predicting the aggregate adjustable capacity of the virtual power plant, thereby improving the model's representation ability and prediction accuracy.

[0075] After the attention mechanism is introduced into the CNN-BiLSTM model, the BiLSTM hidden state sequence h = {h1,h2,…,h t} and x={x1,x2,…,x t}, the attention weight is calculated by the following formula

[0076] e ij =tanh(ω h h i +ω s x j +b c )

[0077]

[0078] In the formula, e ij represents the hidden state h of BiLSTM i and CNN output x j The attention score between ij represents the corresponding attention weight, c j is the hidden state of the BiLSTM after attention weighting, which is used to represent the important information of the input sequence. By introducing the attention mechanism, the CNN-BiLSTM model can flexibly adjust the degree of attention to the input sequence, combining the hidden state of the BiLSTM and the CNN output sequence, using dynamic feature fusion and weight adjustment instead of simply superimposing or splicing the two. This combination reduces the focus on irrelevant features in the model, thereby reducing the risk of overfitting and improving the generalization ability of the model. In this way, important information in the data can be captured more accurately, providing reliable support for the prediction of the adjustable capacity of the virtual power plant.

[0079] The CNN-BiLSTM-QR probability prediction model is as follows:

[0080] Quantile regression is a regression analysis method commonly used in statistics. When the specific parameters of the conditional distribution cannot be obtained, quantile regression is more advantageous than the traditional least squares regression model in dealing with the diversity and unevenness of the distribution of explanatory variables, and can fully and accurately describe the possible range of the dependent variable. In the probability prediction of load adjustable capacity, quantile regression can effectively show the impact of the explanatory variables of adjustable capacity at different quantiles on the prediction results of the target adjustable capacity. Its basic expression can be expressed as:

[0081] Q τ (Y|X)=β0(τ)+β1(τ)x1+β2(τ)x2+…β j (τ)x j +ε(τ)

[0082] In the formula, Q τ represents the estimated value of the response variable Y at the quantile τ given the explanatory variable X, τ represents the quantile in the range (0,1), ε(τ) represents the error term at the quantile τ, β j is the quantile regression coefficient. The degree to which the response variable is affected by the explanatory variable at different quantiles can be expressed by the quantile regression coefficient. The values ​​of the regression coefficient at different quantiles are also different. It is usually obtained by minimizing the quantile loss function. The optimization problem can be described by the following formula:

[0083]

[0084] In the formula, the first term is the quantile loss function within the quantile interval, the second term is the quantile loss function outside the quantile interval, and the third term is the penalty function, i = 1, 2, 3, ..., N, N is the number of training samples, and λ is the penalty parameter. In the CNN-BiLSTM model, the present invention applies the quantile regression model to the loss function layer of the neural network to predict the value of the target adjustable capacity variable at different quantile points.

[0085] The loss function is the mean square loss function, and its formula is:

[0086]

[0087] In the formula, is the predicted value of the neural network.

[0088] The CNN-BiLSTM-QR model after adding quantile regression is shown as follows:

[0089] Q τ (Y|X)=f(X,W(τ),V(τ))

[0090] Where W(τ) and V(τ) are the weight matrices between the hidden layer and the input and output layers of the neural network, respectively.

[0091] Substituting the formula minimizing the quantile loss function into the above CNN-BiLSTM-QR model can obtain the quantile loss function QLOSS and the optimization target of the model, as shown in the following formula:

[0092]

[0093] The quantile loss function is an indicator used to measure the difference between the model prediction results and the actual observation values. The neural network quantile regression model after changing the mean square loss function to the quantile loss function can not only predict the central trend of the virtual power plant aggregate adjustable capacity variable by minimizing the quantile loss function, but also provide prediction intervals at different confidence levels, thereby more comprehensively describing the distribution characteristics and potential risks of the target variable. When τ = 0.5, the quantile loss function value is equal to the original loss function and can be output as a point prediction result.

[0094] However, in the prediction process using the CNN-BiLSTM-QR model, due to the influence of uncertainty in the input data, such as future weather data has a certain uncertainty rather than a fixed value, the adjustable capacity at various quantile points may have a certain deviation, and even overlap of quantile results may occur. The temperature and user baseline load set in the present invention have uncertainty, and each obeys a normal distribution with the respective parameters in the prediction set as the mean. In order to solve this problem, the present invention uses the Monte Carlo method to sample and generate M prediction scenarios, and the probability prediction model generates the estimated value at each quantile point. The process is as follows:

[0095] The CNN-BiLSTM-QR probability prediction model is generated by training the test set and validation set.

[0096] The Monte Carlo method is used to sample from the test set to generate samples with a length equal to the prediction target, and this is repeated m times to obtain the sample set M.

[0097] The CNN-BiLSTM-QR probability prediction model is used to predict the sample set M, and a total of mγ adjustable capacity prediction sequences under γ quantiles are obtained.

[0098] Summarize and calculate the distribution of all sample prediction results and reconstruct the adjustable capability prediction value at the required quantile point.

[0099] Finally, the model uses the kernel density estimation probability density function to eliminate the assumption of the prior distribution of the adjustable potential prediction results. Kernel density estimation is a non-parametric estimation method that better expresses the probability distribution characteristics of adjustable potential with a continuous probability density function. The formula is as follows:

[0100]

[0101] In the formula, is at the quantile τ i The adjustable capacity quantile estimate under γ represents the total number of quantiles, h is the window width, which is determined by empirical method, K(·) is the kernel function, and the present invention adopts Gaussian kernel function:

[0102]

[0103] The probability prediction process of aggregated adjustable capacity of virtual power plants based on CNN-LSTM-QR is as follows:

[0104] Dataset collection: Based on the comprehensive evaluation results of the adjustable potential of the virtual power plant, the baseline load of users in each node area, the number of users of each type in the node, the maximum allowable delay, electricity price, time information and temperature data are collected as the input set, and the aggregated adjustable capacity of the virtual power plant is used as the output set, which is divided into training set, validation set and test set.

[0105] Data preprocessing: Before prediction, the measured data needs to be tested for outliers, filled with missing values, and normalized.

[0106] Correlation analysis: Perform correlation analysis on the data set based on the maximum information coefficient and reconstruct the feature input model with high correlation.

[0107] Construct a probability prediction model: Build a CNN-LSTM model, LSTM model, Stack-LSTM model, CNN model, MLP model and XGBoost model for comparison, set model hyperparameters, input the test set and validation set into the prediction model, use the test set to train the network parameters of the prediction model, and regularly use the validation set for cross-validation during training to monitor the model training. Use quantile regression for probability prediction, take 19 quantile points τ∈[0.05,0.1,…,0.95], and there are 19 loss functions in the corresponding probability prediction model.

[0108] Test set prediction: The test set is input into the trained prediction model to obtain the prediction results of the aggregated adjustable capacity of the virtual power plant at 19 quantiles, and the output when the quantile τ = 0.5 is taken as the point prediction result.

[0109] Construct prediction intervals: Select τ = [0.05, 0.95], τ = [0.1, 0.9], …, τ = [0.45, 55] to construct 9 prediction intervals with confidence levels ranging from 90% to 10% respectively.

[0110] (4) Use evaluation indicators to evaluate the performance of point prediction results and probability prediction results.

[0111] The prediction result performance evaluation process is as follows:

[0112] The performance evaluation process requires the use of point prediction evaluation indicators and confidence interval evaluation indicators. The evaluation indicators of point prediction include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). The mean absolute error is used to evaluate the overall quality of the prediction results, the root mean square error can reflect the overall fluctuation of the prediction results, and the mean absolute percentage error can reflect the prediction performance. The formulas of the three evaluation indicators are as follows:

[0113]

[0114] Where d is the dimension of the output time series data, and f i are the predicted value and actual value of the adjustable capacity at time i respectively.

[0115] A confidence interval is a range used in statistics to express parameter estimation, which includes the possibility of the true parameter value. The confidence level refers to the degree of confidence in the confidence interval. The confidence interval evaluation index evaluates the quality, accuracy and coverage of the confidence interval to determine its accuracy and credibility for parameter estimation. The confidence interval evaluation index used in the present invention includes interval coverage (Prediction Interval Coverage Probability, PICP), interval average bandwidth (Prediction Interval Normalized Average Width, PINAW), and interval score (Interval Score, IS).

[0116] Interval coverage. Interval coverage is an important indicator to measure the reliability of interval prediction, reflecting the probability that the actual value falls within the confidence interval. The closer its value is to the confidence level, the more reliable the interval is. Given the confidence level of the prediction interval, its definition is as follows:

[0117]

[0118] Where L i and U i are the upper and lower limits of the prediction interval respectively.

[0119] Interval average bandwidth. When the interval is wide enough, the interval coverage rate can easily reach 100%. In this case, the interval coverage rate becomes meaningless because the interval value is too conservative. The interval average bandwidth is an indicator of the sensitivity of the observation interval, reflecting the discreteness of the predicted value within the confidence interval. It is defined as follows:

[0120]

[0121] The interval average bandwidth reflects the uncertainty contained in the confidence interval. The smaller its value, the more sensitive the confidence interval.

[0122] Interval score. Interval coverage and interval average bandwidth reflect the reliability and sensitivity of the prediction interval respectively, but calculating one of them alone cannot evaluate the overall performance of the prediction interval. The interval score index is an evaluation index that comprehensively considers coverage and interval width. It is defined as follows:

[0123]

[0124] η=U i -L i

[0125] Where α is the confidence level, η is the interval width, and a certain penalty value is given when the actual value does not fall within the confidence interval. The interval score is used as a comprehensive evaluation indicator of probability prediction. The smaller its value is, the better the comprehensive performance of the prediction interval is.

[0126] The computer storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the above-mentioned method for predicting the probability of aggregated adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism.

[0127] The computer device described in the present invention includes a storage, a processor, and a computer program stored in the storage and executable on the processor. When the processor executes the computer program, the above-mentioned method for predicting the probability of aggregated adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism is implemented.

[0128] In order to verify the rationality and effectiveness of the proposed method, the present invention uses the historical load data of 3,300 users in a certain area as a data set, including 40,992 sets of data with a sampling interval of 15 minutes from May 30, 2018 to July 31, 2019. Each set of data contains the user baseline load of each regional node at that moment, the number of users of each type in the region, the maximum allowed delay, electricity price, time data and temperature data, and the aggregate adjustable capacity evaluation results of the virtual power plant, including the aggregate adjustable capacity of the virtual power plant and the aggregate adjustable capacity data of each regional node at that moment. The data set is divided into training set, validation set and test set according to the time period.

[0129] like Figure 3 As shown in the figure, the calculation results of the mutual correlation between regional nodes and edge nodes and the correlation between regional nodes are analyzed using the maximum information coefficient. The lighter the color, the higher the correlation. The correlation between regional nodes and edge nodes reflects the impact of the load under each regional node on the aggregation and adjustable capacity of edge nodes. The correlation heat map between nodes reflects the similarity of input data between nodes, among which there are obvious partitions.

[0130] In order to verify the rationality and effectiveness of the proposed method, the present invention uses CNN-BiLSTM, CNN-BiLSTM, CNN, Stack-LSTM, MLP, LSTM and XGBoost algorithms combined with cloud-edge mechanism for comparison. The prediction errors are shown in Table 1. It can be seen that the CNN-BiLSTM algorithm combined with cloud-edge mechanism proposed in the present invention has the best performance among the point prediction algorithms. Taking the prediction results of the first day in the prediction set as an example, the prediction results of the aggregate adjustable capacity of the virtual power plant during the peak period under various algorithms are as follows Figure 4 and Figure 5 As shown in the figure, the predicted values ​​of aggregate peak shaving and aggregate valley filling capacities when the maximum allowable delay is 6s. The CNN-BiLSTM algorithm combined with the cloud-edge mechanism at peak times can generally better fit the actual adjustable capacity curve.

[0131] Figure 6 The quantiles are 19 conditional quantile prediction values ​​with quantiles ranging from 0.05 to 0.95, covering 10 confidence intervals with confidence levels ranging from 10% to 90%. The quantiles are point prediction results without considering uncertainty. The higher the confidence level, the wider the interval. The confidence interval results when the quantiles are 0.05 to 0.95 are as follows: Figure 7 As shown in the figure, the confidence level is 90% at this time. Based on the Monte Carlo sampling method, the upper and lower limits of the actual interval when the confidence level is 90% are obtained, and the probability prediction result is close to the actual interval. When the load is at a low point, it is closer to the actual interval, and there is a slightly smaller prediction value at the peak load. The prediction performance of each probability prediction algorithm using quantile regression is shown in Table 2. When the confidence levels are 90%, 80%, and 70%, respectively, the interval coverage of the CNN-BiLSTM algorithm combined with the cloud-edge mechanism is closer to the confidence level and has better reliability. At the same time, the three methods are slightly insufficient in predicting the upper limit value; the proposed method is basically the smallest in the interval average bandwidth index, so it has a certain degree of sensitivity while ensuring reliability; the interval score reflects that the proposed method has the best comprehensive performance index.

[0132] As shown in Table 3, the training time of each algorithm model is faster than that of other algorithms that directly input all data into a single network. Therefore, the cloud-edge mechanism not only improves the accuracy of prediction, but also has high efficiency.

[0133] After obtaining the prediction results at each quantile at each time, the probability density curve prediction results at 0:00, 6:00, 12:00 and 18:00 are selected. The prediction results of the quantiles in each time period are used as input, and the probability distribution of the aggregated peak shaving capacity is drawn based on the Gaussian kernel density estimation. Figure 8 The actual point prediction values ​​at the four moments are all near the highest point of the probability density curve, which further verifies the reliability of the probability prediction results.

[0134] Table 1

[0135]

[0136] Table 2

[0137]

[0138] Table 3

[0139]

Claims

1. A method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism, characterized in that: The following steps are involved: (1) Build a cloud-edge collaborative network system for regulating virtual power plants; (2) Preprocess the collected edge node data, select features for each edge node using the maximum information coefficient method, and obtain a data set as the input of the corresponding edge node in the cloud-edge collaborative network system. The edge node performs regional adjustable capacity evaluation; (3) Constructing a CNN-BiLSTM-QR model, which is used to input the input data set of each edge node obtained from the cloud and the output regional adjustable capacity evaluation value to obtain the predicted value of the aggregate adjustable capacity of the virtual power plant; wherein the CNN-BiLSTM-QR model includes a local CNN layer, a local pooling layer, a global CNN layer, a global pooling layer, a BiLSTM layer, a fully connected layer and an output layer; the local CNN layer is used to input the input data set of the edge node and the output regional adjustable capacity evaluation value to obtain the data features of each edge node and the region described therein; The global CNN layer is used to input the data features of the edge nodes and the areas described by all local pooling layers to obtain the global data features of the cloud layer. The loss function of the BILSTM neural network adopts the quantile loss function, sets multiple quantiles and predicts the adjustable capacity at each quantile. The data features are then mapped to the output layer with the same number of quantiles through the fully connected layer as the output result. The output at a specific quantile whose quantile loss function value is equal to the original loss function is selected as the point prediction result output. By constructing prediction intervals with different confidence levels in the quantile sequence, the output result of the interval is output as the probability interval prediction result.

2. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 1 is characterized in that: The edge node data collected include user baseline load, number of each type of users in the node, maximum allowable delay, electricity price, date, real-time temperature and daily maximum temperature.

3. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 1 is characterized in that: The BiLSTM layer in step (3) introduces the attention mechanism. First, the BiLSTM hidden state sequence h = {h1,h2,…,h t } and x={x1,x2,…,x t }, attention weight α ij for e ij =tanh(ω h h i +ω s x j +b c ) Among them, e ij and e ik represents the hidden state h of BiLSTM i And the global pooling layer output x j The attention score between j Represents the hidden state of the attention-weighted BiLSTM, which is used to represent the important information of the input sequence.

4. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 3 is characterized in that: The quantile loss function QLOSS in step (3) is: Where τ represents the quantile, Y i represents the actual adjustable capacity of the virtual power plant, f(·) represents the solution function of the estimated value of the response variable at the quantile τ under the condition of given characteristic explanatory variables, X represents the given characteristic explanatory variables, W(t) and V(t) represent the weight matrices between the hidden layer and the input and output layers of the neural network, respectively, ω i and v i Respectively represent the i-th element in the W(t) and V(t) matrices, and l represents the penalty parameter; Set multiple quantiles and predict the adjustable capacity at different quantiles separately, and map various data features to the output layer consistent with the number of quantiles through the fully connected layer as the output result. Select the output at a specific quantile whose quantile loss function value is equal to the original loss function as the point prediction result output. Construct prediction intervals with different confidence levels by boxing in the quantile sequence, and output the output result of the interval as the probability interval prediction result; The result output Q of the CNN-BiLSTM-QR model τ for Q τ (Y|X)=f(X,W(τ),V(τ))。 5. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 4 is characterized in that: The Monte Carlo sampling method is used to generate samples of M prediction scenarios, which are input into the trained CNN-BiLSTM-QR model to obtain the predicted value of the aggregated adjustable capacity of the virtual power plant. The predicted value of the aggregated adjustable capacity of the virtual power plant is corrected through the kernel density estimation probability density function to obtain the final predicted value of the aggregated adjustable capacity of the virtual power plant.

6. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 5 is characterized in that: The probability density function for in, Indicates that at the quantile τ i The adjustable capacity quantile estimate under , γ represents the number of quantiles, h is the window width, and K(·) represents the Gaussian kernel function.

7. The method for predicting the aggregate adjustable capacity of a virtual power plant based on a cloud-edge collaboration mechanism according to claim 6 is characterized in that: Also includes (4) The performance of the final prediction results is evaluated through point prediction evaluation indicators and confidence interval evaluation indicators. The evaluation indicators of point prediction include mean absolute error, root mean square error and mean absolute percentage error.

8. A prediction system for aggregated adjustable capacity of virtual power plants based on cloud-edge collaboration mechanism, characterized in that: include System building module, used to build a cloud-edge collaborative network system for regulating virtual power plants; The data collection and processing module is used to collect edge node data and perform preprocessing. It selects features for each edge node through the maximum information coefficient method to obtain a data set as the input of the corresponding edge node in the cloud-edge collaborative network system. The edge node performs regional adjustable capacity evaluation; A model building module is used to build a CNN-BiLSTM-QR model, which is used to input the input data set of each edge node obtained from the cloud and the output regional adjustable capacity evaluation value to obtain the predicted value of the aggregated adjustable capacity of the virtual power plant, wherein the CNN-BiLSTM-QR model includes a local CNN layer, a local pooling layer, a global CNN layer, a global pooling layer, a BiLSTM layer, a fully connected layer and an output layer; the local CNN layer is used to input the input data set of the edge node and the output regional adjustable capacity evaluation value to obtain the data features of each edge node and the region described therein; The global CNN layer is used to input the data features of the edge nodes and the areas described by all local pooling layers to obtain the global data features of the cloud layer. The loss function of the BILSTM neural network adopts the quantile loss function, sets multiple quantiles, predicts the adjustable capabilities at different quantiles separately, and maps various data features to the output layer consistent with the number of quantiles through the fully connected layer as the output result. The output at a specific quantile whose quantile loss function value is equal to the original loss function is selected as the point prediction result output. By selecting prediction intervals of different confidence levels in the quantile sequence, the output result of the interval is output as the probability interval prediction result.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images