Short-term region net load prediction method

Abstract

The invention is suitable for the technical field of load prediction, and provides a short-term region net load prediction method, which comprises the following steps: acquiring power data and weather data of a distributed photovoltaic development region; based on the power data and the weather data, utilizing a plurality of preset prediction models to perform multi-type prediction of distributed photovoltaic output and load, and obtaining an initial prediction result of each preset prediction model; based on the initial prediction result of each preset prediction model, evaluating the credibility of each preset prediction model from the difference and relevance of the prediction results; and based on the credibility of each preset prediction model, performing dynamic weighted combination on the initial prediction result of each preset prediction model to obtain a net load prediction result. According to the method, the weight of each model can be dynamically adjusted according to the current task and scene by utilizing the weighted combination of the dynamic weights of the models, so that a more accurate net load prediction result is obtained, and the prediction precision is improved by 1.44-1.82%.

Description

Technical Field

[0001] The present application belongs to the technical field of load forecasting, and in particular, relates to a short-term regional net load forecasting method. Background Art

[0002] There are two main ways to predict net load: direct prediction and indirect prediction. Indirect prediction refers to predicting distributed energy and power load separately, solving the difference between the two to predict net load; direct prediction refers to exploring the fluctuation trend of net load data, and then predicting net load.

[0003] At present, the research methods for net load forecasting at home and abroad mainly include traditional methods and artificial intelligence methods. Traditional net load forecasting methods mainly include time series method, regression analysis method and other statistical methods. Traditional net load forecasting methods can better reflect the law of net load changes, but they have the disadvantages of complex modeling process or high data quality requirements. And traditional net load forecasting methods often have some limitations when facing the complexity and uncertainty of modern power systems. First, with the development of smart grids, a large amount of nonlinear power load data has accumulated in the power grid, such as load data driven by the increase of distributed generation and the development of energy storage technology. Traditional net load forecasting methods may show great limitations when processing these nonlinear data. Secondly, in order to facilitate calculation and analysis, traditional methods often simplify complex load changes. For example, the unit consumption method only considers the relationship between product output and unit electricity consumption, while ignoring other possible influencing factors. This simplification may cause the model to fail to accurately reflect the complexity and diversity of actual loads.

[0004] With the continuous development of artificial intelligence, more and more scholars have introduced artificial intelligence technology to predict net load. However, how to use artificial intelligence methods to achieve accurate prediction needs to be studied urgently. Summary of the invention

[0005] The embodiment of the present application provides a short-term regional net load forecasting method, which can achieve accurate forecasting using artificial intelligence methods.

[0006] This application is implemented through the following technical solutions:

[0007] In a first aspect, an embodiment of the present application provides a short-term regional net load forecasting method, comprising:

[0008] Obtain power and weather data in distributed photovoltaic development areas;

[0009] Based on power data and weather data, multiple preset prediction models are used to perform multi-type predictions of distributed photovoltaic output and load, and the initial prediction results of each preset prediction model are obtained;

[0010] Based on the initial prediction results of each preset prediction model, the credibility of each preset prediction model is evaluated from the difference and correlation of the prediction results;

[0011] Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the net load prediction result.

[0012] In one possible implementation, the credibility includes uncertainty and relevance;

[0013] Based on the initial prediction results of each preset prediction model, the credibility of each preset prediction model is evaluated from the difference and correlation of the prediction results, including:

[0014] Based on the evaluation method of information entropy uncertainty, the time series information entropy of the initial prediction results of each preset prediction model is calculated to obtain the uncertainty of the initial prediction results of each preset prediction model;

[0015] Based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated.

[0016] In a possible implementation, the information entropy H(X) corresponding to the uncertainty of the initial prediction results of each preset prediction model is expressed as:

[0017]

[0018] Among them, p(x i ) is the initial prediction result of the preset prediction model X is x i The probability of; i is the value number of the prediction result X, n is the number of the value numbers of the prediction result X.

[0019] In a possible implementation, based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated, including:

[0020] Normalizing the initial prediction results of each preset prediction model to obtain a processed data set;

[0021] Based on the prediction accuracy evaluation index, the processed data set corresponding to the initial prediction result with the highest prediction accuracy is selected as the reference sequence, and the initial prediction results of other preset prediction models are used as the comparison sequence;

[0022] The similarity between the reference sequence and the comparison sequence is calculated to determine the correlation between the initial prediction results of each preset prediction model; the higher the correlation, the closer the relationship between the two sequences.

[0023] In a possible implementation, based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain a net load prediction result, including:

[0024] Based on the uncertainty of the initial prediction results of each preset prediction model, a weight value is assigned to the initial prediction results of each preset prediction model to obtain the moment-by-moment weight value of the prediction results of each model;

[0025] Based on the moment-by-moment weight values ​​of the prediction results of each model, a combined prediction result is obtained;

[0026] Based on the correlation between the combined prediction result and the initial prediction results of each preset prediction model, the accuracy of the combined prediction result is verified to obtain the net load prediction result.

[0027] In a possible implementation, based on the uncertainty of the initial prediction results of each preset prediction model, a weight value is assigned to the initial prediction results of each preset prediction model, including:

[0028] According to the uncertainty of the initial prediction results of each preset prediction model from low to high, the initial prediction results of each preset prediction model are assigned weight values ​​from high to low.

[0029] In a possible implementation, the multiple preset prediction models include a deep neural network, a long short-term memory neural network, a Transformer model, XGBoost, and a support vector machine.

[0030] In a possible implementation manner, the short-term regional net load forecasting method further includes:

[0031] Continuously obtain real-time power data and real-time weather data in distributed photovoltaic development areas, and update the dynamic weights of various preset prediction models based on real-time power data and real-time weather data.

[0032] In a possible implementation, while dynamically weighting and combining the initial prediction results of each preset prediction model based on the credibility of each preset prediction model to obtain the net load prediction result, the short-term regional net load prediction method also includes:

[0033] Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the baseline load and distributed photovoltaic power generation power.

[0034] In a possible implementation, the power data includes grid-side net load data, user-side photovoltaic output, and baseline load data;

[0035] Weather data includes weather forecast values ​​for a period of time in the future.

[0036] Compared with the related art, the embodiments of the present application have the following beneficial effects:

[0037] The short-term regional net load forecasting method of the embodiment of the present application, in view of the different adaptability of different forecasting models in various load and photovoltaic power generation scenarios, utilizes the weighted combination of model dynamic weights to dynamically adjust the weights of each model according to the current task and scenario, thereby obtaining a more accurate net load forecasting result. Using the forecasting method of artificial intelligence, it is possible to explore the nonlinear relationship between complex influencing factors and net load, evaluate the credibility of each forecasting model from the two dimensions of difference and correlation, and explore the changing trend of net load under time series. It can more comprehensively consider various influencing factors and reduce the errors that may be caused by a single model, thereby improving the accuracy and reliability of the forecast. The forecasting accuracy of this method is improved by 1.44% to 1.82%, thereby providing strong support for the stable operation and optimized management of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 This is an application background diagram of a short-term regional net load forecasting method provided by an embodiment of the present application;

[0040] Figure 2 It is a flow chart of a short-term regional net load forecasting method provided by an embodiment of the present application;

[0041] Figure 3 It is a schematic diagram of the overall process of constructing baseline load, distributed photovoltaic power generation power and net load provided by an embodiment of the present application;

[0042] Figure 4 It is a structural diagram of a Transformer model provided in an embodiment of the present application;

[0043] Figure 5 It is a schematic diagram of the structure of XGBoost provided in one embodiment of the present application;

[0044] Figure 6 is a schematic diagram of a flow chart for determining uncertainty and correlation provided by an embodiment of the present application;

[0045] Figure 7 It is a schematic diagram of the baseline load forecasting result of the preset forecasting model provided in one embodiment of the present application and its information entropy and representativeness;

[0046] Figure 8 It is a schematic diagram of photovoltaic power prediction results and information entropy and representativeness of a preset prediction model provided in an embodiment of the present application;

[0047] Fig. 9 is a baseline load forecast result and model weight provided by an embodiment of the present application;

[0048] Fig.10 The photovoltaic power generation prediction result and model weight provided in an embodiment of the present application;

[0049] Fig.11 is a net load prediction result provided by an embodiment of the present application;

[0050] Fig.12 It is a structural schematic diagram of a short-term regional net load forecasting device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0051] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0052] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0053] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0054] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0055] References to "an embodiment", "one embodiment" or "some embodiments" described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0056] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0057] Reference Figure 1 1 is an application scenario diagram of the short-term regional net load forecasting method of the present invention. The forecasting method runs in the server 102. The server 102 obtains power data and related data from the power grid database 101.

[0058] Figure 2 is a flow chart of a short-term regional net load forecasting method provided by an embodiment of the present application, referring to Figure 2 , the short-term regional net load forecasting method includes:

[0059] Step 201, obtaining power data and weather data in a distributed photovoltaic development area.

[0060] Exemplarily, the power data may include grid-side net load data, user-side photovoltaic output and baseline load data; the weather data may include weather forecast values ​​for a period of time in the future.

[0061] Step 202 , based on the power data and weather data, using multiple preset prediction models, perform multi-type prediction of distributed photovoltaic output and load, and obtain initial prediction results of each preset prediction model.

[0062] Exemplarily, the multiple preset prediction models may include artificial intelligence models such as deep neural networks, long short-term memory neural networks, Transformer models, XGBoost and support vector machines.

[0063] Step 203 , based on the initial prediction results of each preset prediction model, the credibility of each preset prediction model is evaluated from the difference and correlation of the prediction results.

[0064] Step 204 , based on the credibility of each preset prediction model, dynamically weighted combination is performed on the initial prediction results of each preset prediction model to obtain a net load prediction result.

[0065] For example, see Figure 3 Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the net load prediction result, including:

[0066] Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the baseline load and distributed photovoltaic power generation power.

[0067] In this embodiment, in view of the different adaptability of different prediction models in various load and photovoltaic power generation scenarios, a weighted combination of model dynamic weights is used to dynamically adjust the weights of each model according to the current task and scenario, thereby obtaining a more accurate net load prediction result. The prediction method using artificial intelligence can explore the nonlinear relationship between complex influencing factors and net load, evaluate the credibility of each prediction model from the two dimensions of difference and correlation, and explore the changing trend of net load in time series, which can more comprehensively consider various influencing factors and reduce the errors that may be caused by a single model, thereby improving the accuracy and reliability of the prediction.

[0068] In one embodiment, in step 202, the plurality of preset prediction models may include artificial intelligence models such as deep neural networks, long short-term memory neural networks, Transformer models, XGBoost and support vector machines.

[0069] For example, a deep neural network (DNN) is a neural network with multiple hidden layers. DNN includes an input layer, a hidden layer, and an output layer, and the layers are connected in a fully connected manner. Assuming that the DNN model has i neurons in the l-1 layer and j neurons in the l layer, the calculation method of the neuron output is shown in formula (1). DNN has a strong feature expression ability and can extract multi-level features from the original data, which has significant advantages in processing complex data. At the same time, DNN has a strong generalization ability by learning the inherent fluctuation laws and patterns of the data.

[0070]

[0071] in, is the weight from the kth neuron in the l-1th layer to the i-th neuron in the lth layer; a is the neuron output result; b is the offset vector.

[0072] For example, the long short-term memory (LSTM) neural network is a time-trained network designed to solve long-term problems. Compared with the traditional recurrent neural network RNN, LSTM has a stronger memory capacity by designing a gating structure to handle long-term dependencies. The basic unit of LSTM contains an input gate, an output gate, and a forget gate, which together control the state update of the memory unit. Through reasonable weight distribution and gating mechanism, LSTM can effectively handle long-term dependencies in sequence data.

[0073] The forget gate in LSTM determines what information will be forgotten. The input h t-1 and x t , outputs a number between 0 and 1 in the neuron state, where 1 represents complete retention and 0 represents complete forgetting. In this way, LSTM can memorize important information for a long time and can dynamically adjust with the input, as shown in formula (2). The cell state in LSTM is the hidden state of the neuron, which is the "memory" of the recurrent neural network for the input data. It is used to represent the "memory" of the neuron after time t, as shown in formula (3). The input gate and output gate in LSTM determine the input and output values. The input gate saves the neuron state value, and the output gate determines that the neuron state needs to be output, as shown in formulas (4) and (5).

[0074] f t =σ(W f ·[h t-1 ,x t ]+b f ] (2)

[0075] C t =tanh(W c ·[h t-1 ,x t ]+b c ) (3)

[0076] i t =σ(W i ·[h t-1 ,x t ]+b i ) (4)

[0077] o t =σ(W o ·[h t-1 ,x t ]+b o ) (5)

[0078] Among them, f t 、i t , C t and t are the activation function output of the forget gate, the activation function output of the input gate, the activation function output of the cell state, and the activation function output of the output gate; W f , W i , W c and W o are the feedforward network weight matrix of the forget gate, the feedforward network weight matrix of the input gate, the feedforward network weight matrix of the cell state, and the feedforward network weight matrix of the output gate; b f 、b i 、b c and b o are the feedforward network bias vector of the forget gate, the feedforward network bias vector of the input gate, the feedforward network bias vector of the cell state, and the feedforward network bias vector of the output gate; h t-1 is the hidden state at the previous moment.

[0079] For example, the Transformer model is a deep learning model. Its core idea is to use self-attention mechanism and multi-head attention mechanism to process input data, and use structures such as position feedforward network and residual connection to improve the performance of the model. The model structure is as follows Figure 4 Compared with the traditional RNN model, the Transformer model has better parallel computing capabilities and stronger expressiveness.

[0080] The calculation method of the self-attention mechanism in the Transformer model is shown in formula (6). Given an input sequence X, the linear transformation W q , W k and W v The query vector Q, key vector K and value vector V are obtained respectively. Then, the query vector Q and key vector K are used to perform a dot product operation, and then the weight is obtained through the softmax function. Finally, the weight is multiplied by the value vector V to obtain the output of the self-attention. The calculation method of the self-fully connected feedforward network in the Transformer model is shown in formula (7):

[0081]

[0082] FFN(x)=max(0,xW1+b1)W2+b2 (7)

[0083] Among them, Attention() is the attention mechanism; FFN(x) is the feedforward neural network; Q is the query vector; K is the key vector; V is the value vector; d kis the number of columns of the Q and K matrices, i.e., the vector dimension; x is the input of the connection layer; W1 and W2 are weight matrices; b1 and b2 are offset vector matrices.

[0084] For example, XGBoost is a machine learning model based on the gradient boosting decision tree algorithm, which is efficient, scalable and flexible. XGBoost uses the gradient boosting algorithm to fit the data by iteratively building a series of decision trees. The model structure is as follows Figure 5 shown.

[0085] The objective function of XGBoost is shown in formulas (8) and (9), which consists of a loss function and a regularization function. XGBoost uses a second-order Taylor expansion to approximate the loss function and uses the gradient descent method to optimize the model parameters. This optimization method can converge quickly and achieve better performance. In each iteration, the model tries to minimize the loss function, thereby continuously optimizing the model's prediction performance. XGBoost is a powerful, efficient, and flexible machine learning model. Its gradient boosting algorithm and decision tree structure enable it to perform well in processing complex data sets, providing accurate prediction results and stable performance.

[0086]

[0087] Where L(φ) is the expression in linear space; i is the i-th sample, k is the k-th tree, k = 1, 2, ..., K; is the i-th sample x i The predicted value of . is the loss function; Ω(f) is the regularization term.

[0088] For example, support vector machine (SVM) is a supervised learning model used for classification and regression analysis. SVM is based on statistical learning theory and achieves classification by finding a decision boundary that can maximize the separation of data points of different categories. The goal of SVM is to find a hyperplane that can correctly classify the training data and is as far away from the training data as possible to improve generalization ability, as shown in formula (10). Support vectors are the training data points closest to the hyperplane, which determine the position of the hyperplane, as shown in formula (11). During the training process, SVM finds the optimal hyperplane through optimization algorithms (such as quadratic programming) so that it can maximize the classification interval and minimize the error. Support vectors play a key role in the training process because they determine the complexity and generalization ability of the model.

[0089] w·x+c=0 (10)

[0090] y i (w·x i+c)=0 (11)

[0091] Among them, w is the normal vector; x is the input feature; c is the intercept; y is the i is the data point x i The category label is 1 or -1.

[0092] This embodiment combines a variety of mainstream prediction models, and can make full use of their respective advantages to improve the prediction accuracy of photovoltaic output and load. For example, the fitting ability of FCNN and the global feature extraction ability of Transformer can be used, combined with the generalization performance of SVM, to build a more accurate and stable prediction model. Multi-type modeling prediction can reduce the prediction risk of a single model under specific conditions and improve the robustness of the entire prediction system. When a model performs poorly due to data anomalies or environmental changes, other models can play a complementary role to ensure the stability and reliability of the prediction results.

[0093] One embodiment, participate Figure 6 , credibility includes uncertainty and relevance. Step 203 includes:

[0094] Step 2031, based on the evaluation method of information entropy uncertainty, calculate the time series information entropy of the initial prediction results of each preset prediction model to obtain the uncertainty of the initial prediction results of each preset prediction model.

[0095] The information entropy H(X) corresponding to the uncertainty of the initial prediction results of each preset prediction model is expressed as:

[0096]

[0097] Among them, p(x i ) is the initial prediction result of the preset prediction model X is x i The probability of; i is the value number of the prediction result X, n is the number of the value numbers of the prediction result X.

[0098] The information entropy of the prediction results of each model is calculated to characterize the amount of information in the prediction results of each model. Then the information entropy of each moment is calculated moment by moment to characterize the uncertainty of the prediction results at each moment.

[0099] The evaluation method based on information entropy uncertainty is a method for evaluating the uncertainty of time series data. Its basic principle is to use the concept of information entropy to measure the complexity and uncertainty of time series data. The calculation method of information entropy is shown in formula (12), which indicates the degree of chaos of the system and also indicates the degree of knowledge of the amount of information in the system. The time series uncertainty evaluation method based on information entropy has the advantages of simplicity, easy implementation and strong scalability, and can be applied to the processing and analysis of various time series data. At the same time, this method can also be used in combination with other machine learning algorithms to improve the accuracy and reliability of time series data analysis.

[0100] Step 2032, based on the grey correlation analysis method, calculate the correlation of the initial prediction results of each preset prediction model.

[0101] Grey correlation analysis is a statistical method used to measure the degree of correlation between various factors in a system. It determines the degree of correlation between them by comparing the similarities between data sequences. When the uncertainty of data information is large, the grey correlation analysis model can effectively reveal the relationship and influence mechanism between various factors. Based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated, including:

[0102] The initial prediction results of each preset prediction model are normalized to obtain a processed data set, which can eliminate the influence of different numerical ranges.

[0103] Based on the prediction accuracy evaluation index, the processed data set corresponding to the initial prediction result with the highest prediction accuracy is selected as the reference sequence, and the initial prediction results of other preset prediction models are used as the comparison sequence.

[0104] The similarity between the reference sequence and the comparison sequence is calculated to determine the correlation between the initial prediction results of each preset prediction model; the higher the correlation, the closer the relationship between the two sequences.

[0105] By calculating information entropy, this embodiment can identify the part of the prediction result with high uncertainty, and then focus on analyzing and optimizing this part of the data, which helps to improve the accuracy of the overall prediction and reduce errors. By calculating the correlation, it is also possible to identify the prediction model or prediction result with a high correlation with the actual situation. These models or results are usually more reliable and can serve as an important basis for subsequent decision-making.

[0106] In one embodiment, step 204 includes:

[0107] First, based on the uncertainty of the initial prediction results of each preset prediction model, a weight value is assigned to the initial prediction results of each preset prediction model to obtain the moment-by-moment weight value of the prediction results of each model.

[0108] Exemplarily, based on the uncertainty of the initial prediction results of each preset prediction model, the weight value of the initial prediction results of each preset prediction model is assigned, including: according to the uncertainty of the initial prediction results of each preset prediction model from low to high, the weight value of the initial prediction results of each preset prediction model is assigned from high to low. That is, a higher weight is assigned to a model with lower uncertainty, and a lower weight is assigned to a model with higher uncertainty, so that the weight value of the prediction result of each model at each moment can be obtained.

[0109] Then, based on the moment-by-moment weight values ​​of the prediction results of each model, a combined prediction result is obtained.

[0110] Finally, based on the correlation between the combined prediction result and the initial prediction results of each preset prediction model, the accuracy of the combined prediction result is verified to obtain the net load prediction result.

[0111] In one embodiment, the short-term regional net load forecasting method further includes:

[0112] Continuously obtain real-time power data and real-time weather data in distributed photovoltaic development areas, and update the dynamic weights of various preset prediction models based on real-time power data and real-time weather data.

[0113] For example, by acquiring power data and weather data in real time, the prediction model can respond to environmental changes more quickly and adjust the prediction results in a timely manner. This real-time nature makes the prediction closer to the actual situation and reduces the prediction error caused by data lag. Updating the dynamic weights of the prediction model based on real-time data can make the model pay more attention to the key factors currently affecting photovoltaic power generation. For example, on cloudy or rainy days, the model may increase the weight of weather factors such as cloud thickness and rainfall, while on sunny days it will pay more attention to factors such as solar irradiance. This dynamic adjustment helps to improve the accuracy of the prediction.

[0114] In order to verify the effectiveness of the method of the present invention, the following specific example analysis process is given.

[0115] This paper uses the two-year historical measured data of a distributed photovoltaic development pilot area as an example, with a time resolution of 15 minutes. It includes grid-side net load data, user-side photovoltaic output, baseline load data, and numerical weather forecasts for the next three days. After abnormal data identification and correction, the above data are used for subsequent experiments through cross-validation, in which 75% of the data is divided into a training set, 15% of the data is divided into a test set, and 10% of the data is divided into a validation set. The training set is used to train the model, the validation set is used to select model parameters, and the test set is used to test the final model performance.

[0116] This embodiment uses CNN (M1), Transformer (M2) and LSTM (M3) three types of deep learning models, and XGBoost (M4) and SVM (M5) two types of machine learning models as the benchmark prediction model and control model. Among them, M1-M3 are established by the PyTorch framework, M4 and M5 are established by the Keras framework, and the main parameter settings of each preset prediction model are shown in Table 1.

[0117] Table 1 Main parameter settings of each preset prediction model

[0118]

[0119] This embodiment uses the standardized root mean square error (RMSE) and mean absolute percentage error (MAPE) commonly used in the fields of load forecasting and photovoltaic power generation forecasting as evaluation indicators. The calculation method is as follows:

[0120]

[0121] in, is the predicted value; is the actual value at time t; T is the number of time steps for a prediction result.

[0122] Based on the above experimental data, the time information entropy and grey correlation calculation results of each preset prediction model in the baseline load forecasting and photovoltaic power generation forecasting process of the proposed method are introduced below. The combined prediction results obtained based on the proposed method and the corresponding dynamic combination weight changes are plotted. The above prediction error evaluation index is used to verify the effectiveness of the proposed method in baseline load forecasting and photovoltaic power generation forecasting.

[0123] The calculation results of time information entropy and grey correlation degree of each preset prediction model for baseline load and photovoltaic power generation are as follows: Figure 7 and Figure 8 As shown, it includes the baseline load / photovoltaic power generation output prediction results of each preset prediction model, the representative evaluation index value and model information entropy of each model, the moment-by-moment information entropy of the prediction results and the grey correlation between the model prediction results.

[0124] Depend on Figure 7 and Figure 8It can be seen that the moment-to-moment information entropy of the baseline load forecast results and the photovoltaic power generation forecast results have large changes. The information entropy is large at the location where the forecast results are scattered, and the uncertainty of the forecast results is large; the uncertainty is small at the moment when the forecast results of each model are consistent, and the information entropy decreases accordingly. In terms of the gray correlation of the model forecast results, the similarity between the forecast results of the preset forecast model is not consistent in the two types of tasks. At the same time, there are also cases where multiple forecast results are highly representative.

[0125] Comparing the information entropy and model representativeness of each model, the preset prediction models have great differences in baseline load prediction, the information entropy of M2, M4 and M5 is small, and the representativeness of M2 and M4 models is strong. However, the performance in the photovoltaic power generation prediction results is relatively close, and the information entropy of M2 and M3 is relatively small.

[0126] Based on the above calculation results, we make full use of the advantages of each model and the overall model's ability to estimate the uncertainty of the prediction results, give higher weights to models with lower uncertainty, and give lower weights to models with higher uncertainty. We can get the moment-by-moment weight values ​​of the prediction results of each model, and then get the combined prediction results. The baseline load prediction results and photovoltaic power generation prediction results are shown in Figure 2. Fig. 9 and Fig.10 As shown, the curves of the actual value and the predicted result of the dynamic weight combination model of the proposed method basically coincide.

[0127] Based on the above analysis results, it can be seen that the prediction results of different preset prediction models have different uncertainties, and there is also a certain connection between the prediction results of each model. Effectively utilizing the prediction results of multiple models and performing dynamic weight combination predictions will further improve the prediction accuracy.

[0128] Depend on Fig. 9 It can be seen that compared with the preset prediction model, the proposed method has a stronger fitting effect on the baseline load / power generation value. In particular, at the position where the moment-by-moment information entropy of the prediction results of each preset prediction model is high, there is still a high prediction accuracy.

[0129] Based on the prediction error evaluation index, the error of the proposed method in baseline load and PV power prediction is further calculated, as shown in Table 2. In terms of load prediction, the RMSE of the proposed method is reduced to 1.25MW, and the MAPE is reduced to 4.03%. Compared with the preset prediction model, the RMSE is reduced by 0.518MW, and the MAPE is reduced by 1.13%. In terms of PV power prediction, the RMSE and MAE of the proposed method are 0.83MW and 5.76%, respectively. Compared with the preset prediction model, the RMSE is reduced by 0.296MW, and the MAPE is reduced by 1.44%.

[0130] Table 2 Baseline load and PV power prediction errors of various methods

[0131]

[0132]

[0133] Based on the baseline load and photovoltaic power generation prediction results obtained above, the net load prediction results are synthesized. The prediction results of each method are as follows: Fig.11 The RMSE error and MAPE error index values ​​are shown in Table 3.

[0134] Table 3 Net load prediction error

[0135]

[0136] comprehensive Fig.11 As shown in Table 3, compared with various control methods, the net load fitting effect of the present invention is stronger. RMSE is reduced to 1.06MW, which is 0.41MW lower than the preset prediction model on average; MAPE is reduced to 4.23%, which is 1.82% lower on average.

[0137] In summary, the short-term regional net load forecasting method provided in the embodiment of the present application flexibly integrates different types of forecasting models to fully utilize the advantages of each model. Different models have different advantages in predicting specific types of loads or scenarios. Integration can make up for the shortcomings of a single model. For example, in scenarios with sufficient sunlight and large photovoltaic contributions, the model prefers forecasting algorithms that are good at handling fluctuations in renewable energy, while at night or on cloudy days, the model pays more attention to the accuracy of baseline load forecasts.

[0138] According to the changing trend and characteristics of the baseline load, the weight of each prediction model in the comprehensive prediction is dynamically adjusted to better reflect the actual load situation. According to the adaptability of photovoltaic power generation scenarios, considering the volatility and uncertainty of photovoltaic power generation, and according to the actual output of photovoltaic power generation, the weight of the prediction model is adjusted to improve the prediction accuracy of net load changes and enhance the adaptability of the model.

[0139] In order to effectively adjust the model weights, it is necessary to monitor the power system status in real time, including baseline load, photovoltaic power generation, weather conditions, etc., to obtain the latest data for updating the model weights. At the same time, considering the processing of large amounts of data and real-time requirements, the prediction method of the present invention is efficient and has good scalability when processing large-scale data and complex scenarios.

[0140] The short-term regional net load forecasting method of the present invention has certain advantages in the comprehensive use of multiple forecasting models. In the future, on this basis, we will further study the reliability of net load forecasting in different spatial scales and different typical scenarios, and propose a more accurate forecasting method. This will contribute to promoting the high-proportion photovoltaic power generation and supporting the operation of new power systems.

[0141] See also Fig.12 An embodiment of the present application provides a short-term regional net load forecasting device, which executes the short-term regional net load forecasting method as in the above embodiment, including a data acquisition module 301, a first prediction module 302, a credibility analysis module 303 and a second prediction module 304.

[0142] Data acquisition module 301, used to acquire power data and weather data in the distributed photovoltaic development area;

[0143] The first prediction module 302 is used to perform multi-type prediction of distributed photovoltaic output and load based on power data and weather data using multiple preset prediction models to obtain initial prediction results of each preset prediction model;

[0144] The credibility analysis module 303 is used to evaluate the credibility of each preset prediction model from the difference and correlation of the prediction results based on the initial prediction results of each preset prediction model;

[0145] The second prediction module 304 is used to dynamically weight and combine the initial prediction results of each preset prediction model based on the credibility of each preset prediction model to obtain a net load prediction result.

[0146] Exemplarily, the credibility includes uncertainty and relevance.

[0147] The credibility analysis module 303 is specifically used for:

[0148] Based on the evaluation method of information entropy uncertainty, the time series information entropy of the initial prediction results of each preset prediction model is calculated to obtain the uncertainty of the initial prediction results of each preset prediction model;

[0149] Based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated.

[0150] Exemplarily, the information entropy H(X) corresponding to the uncertainty of the initial prediction results of each preset prediction model is expressed as:

[0151]

[0152] Among them, p(x i ) is the initial prediction result of the preset prediction model X is x iThe probability of; i is the value number of the prediction result X, n is the number of the value numbers of the prediction result X.

[0153] Exemplarily, based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated, including:

[0154] Normalizing the initial prediction results of each preset prediction model to obtain a processed data set;

[0155] Based on the prediction accuracy evaluation index, the processed data set corresponding to the initial prediction result with the highest prediction accuracy is selected as the reference sequence, and the initial prediction results of other preset prediction models are used as the comparison sequence;

[0156] The similarity between the reference sequence and the comparison sequence is calculated to determine the correlation between the initial prediction results of each preset prediction model; the higher the correlation, the closer the relationship between the two sequences.

[0157] Exemplarily, the second prediction module 304 is specifically used for:

[0158] Based on the uncertainty of the initial prediction results of each preset prediction model, a weight value is assigned to the initial prediction results of each preset prediction model to obtain the moment-by-moment weight value of the prediction results of each model;

[0159] Based on the moment-by-moment weight values ​​of the prediction results of each model, a combined prediction result is obtained;

[0160] Based on the correlation between the combined prediction result and the initial prediction results of each preset prediction model, the accuracy of the combined prediction result is verified to obtain the net load prediction result.

[0161] Exemplarily, based on the uncertainty of the initial prediction results of each preset prediction model, the weight value of the initial prediction results of each preset prediction model is assigned, including:

[0162] According to the uncertainty of the initial prediction results of each preset prediction model from low to high, the initial prediction results of each preset prediction model are assigned weight values ​​from high to low.

[0163] Exemplarily, in the first prediction module 302, the multiple preset prediction models include a deep neural network, a long short-term memory neural network, a Transformer model, XGBoost, and a support vector machine.

[0164] Exemplarily, the short-term regional net load forecasting device further includes a data updating module; the data updating module is specifically used for:

[0165] Continuously obtain real-time power data and real-time weather data in distributed photovoltaic development areas, and update the dynamic weights of various preset prediction models based on real-time power data and real-time weather data.

[0166] Exemplarily, based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the net load prediction result. The short-term regional net load prediction device also includes a baseline load and distributed photovoltaic power generation power prediction module; the baseline load and distributed photovoltaic power generation power prediction module is specifically used for:

[0167] Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the baseline load and distributed photovoltaic power generation power.

[0168] Exemplarily, in the data acquisition module 301, the power data includes grid-side net load data, user-side photovoltaic output and baseline load data;

[0169] Weather data includes weather forecast values ​​for a period of time in the future.

[0170] It should be noted that, although several units / modules or sub-units / modules of the short-term regional net load forecasting device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more units / modules described above may be embodied in one unit / module. Conversely, the features and functions of one unit / module described above may be further divided to be embodied by multiple units / modules.

[0171] In addition, although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0172] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the short-term regional net load forecasting method provided in the above-mentioned embodiment of the present application is implemented.

[0173] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0174] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A short-term regional net load forecasting method, characterized in that: include: Obtain power and weather data in distributed photovoltaic development areas; Based on the power data and weather data, using multiple preset prediction models, multi-type predictions of distributed photovoltaic output and load are performed to obtain initial prediction results of each preset prediction model; Based on the initial prediction results of each preset prediction model, the credibility of each preset prediction model is evaluated from the difference and correlation of the prediction results; Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain a net load prediction result.

2. The short-term regional net load forecasting method according to claim 1, characterized in that: The credibility includes uncertainty and relevance; The initial prediction results based on each preset prediction model, respectively evaluating the credibility of each preset prediction model from the difference and correlation of the prediction results, includes: Based on the evaluation method of information entropy uncertainty, the time series information entropy of the initial prediction results of each preset prediction model is calculated to obtain the uncertainty of the initial prediction results of each preset prediction model; Based on the grey correlation analysis method, the correlation of the initial prediction results of each preset prediction model is calculated.

3. The short-term regional net load forecasting method according to claim 2, characterized in that: The information entropy H(X) corresponding to the uncertainty of the initial prediction results of each preset prediction model is expressed as: Among them, p(x i ) is the initial prediction result of the preset prediction model X is x i The probability of; i is the value number of the prediction result X, n is the number of the value numbers of the prediction result X.

4. The short-term regional net load forecasting method according to claim 2, characterized in that: The grey correlation analysis method is used to calculate the correlation of the initial prediction results of each preset prediction model, including: Normalizing the initial prediction results of each of the preset prediction models to obtain a processed data set; Based on the prediction accuracy evaluation index, the processed data set corresponding to the initial prediction result with the highest prediction accuracy is selected as the reference sequence, and the initial prediction results of other preset prediction models are used as the comparison sequence; The similarity between the reference sequence and the comparison sequence is calculated to determine the correlation between the initial prediction results of each preset prediction model; the higher the correlation, the closer the relationship between the two sequences.

5. The short-term regional net load forecasting method according to claim 2, characterized in that: The initial prediction results of the preset prediction models are dynamically weighted and combined based on the credibility of the preset prediction models to obtain a net load prediction result, including: Based on the uncertainty of the initial prediction results of each preset prediction model, a weight value is assigned to the initial prediction results of each preset prediction model to obtain a moment-by-moment weight value of the prediction results of each model; Based on the moment-by-moment weight values ​​of the prediction results of each model, a combined prediction result is obtained; Based on the correlation between the combined prediction result and the initial prediction results of each preset prediction model, the accuracy of the combined prediction result is verified to obtain the net load prediction result.

6. The short-term regional net load forecasting method according to claim 5, characterized in that: The weight values ​​assigned to the initial prediction results of the respective preset prediction models based on the uncertainties of the initial prediction results of the respective preset prediction models include: According to the uncertainty of the initial prediction results of each preset prediction model from low to high, the initial prediction results of each preset prediction model are assigned weight values ​​from high to low.

7. The short-term regional net load forecasting method according to claim 3, characterized in that: The multiple preset prediction models include deep neural network, long short-term memory neural network, Transformer model, XGBoost and support vector machine.

8. The short-term regional net load forecasting method according to claim 1, characterized in that: Also includes: Real-time power data and real-time weather data of the distributed photovoltaic development area are continuously acquired, and the dynamic weights of the various preset prediction models are updated based on the real-time power data and the real-time weather data.

9. The short-term regional net load forecasting method according to claim 1, characterized in that: While dynamically weighting and combining the initial prediction results of the preset prediction models based on the credibility of the preset prediction models to obtain the net load prediction result, the short-term regional net load prediction method also includes: Based on the credibility of each preset prediction model, the initial prediction results of each preset prediction model are dynamically weighted and combined to obtain the baseline load and distributed photovoltaic power generation power.

10. The short-term regional net load forecasting method according to any one of claims 1 to 9, characterized in that: The power data includes grid-side net load data, user-side photovoltaic output and baseline load data; The weather data includes weather forecast values ​​for a period of time in the future.

Images