Distributed new energy power station load prediction method

Abstract

The invention provides a distributed new energy power station load prediction method, and the method comprises the steps: obtaining load prediction data of a new energy power station in a preset time period, preprocessing the load prediction data, and constructing a training data set and a test data set; training the training data set to generate a preliminary load prediction model; evaluating the performance of the preliminary load prediction model through the test data set to generate a preliminary load evaluation result; if the performance requirement is met, inputting a prediction result of the initial load prediction model into an artificial neural network for weighted averaging to obtain a stacked integrated load prediction model; evaluating the performance of the stacked integrated load prediction model on the verification set to generate a load evaluation result; and if the performance requirement is met, optimizing the performance of the stacked integrated load prediction model according to the weather forecast data correction model to output a new energy power station load prediction model so as to solve the problem that the load change and uncertainty of the new energy power station are difficult to accurately predict.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy, in particular to a distributed new energy power station load prediction method. BACKGROUND

[0002] With the rapid development of new energy technology, distributed new energy power stations have become an important trend in the power industry. However, the load prediction problem of new energy power stations has always been a challenge faced by the industry.

[0003] Related load prediction methods are based on historical data and experience on the one hand, which is difficult to accurately predict the load change of new energy power stations; related load prediction methods are based on historical load data and weather data on the other hand, but the accuracy of the prediction is affected by many factors, especially the accuracy of weather forecast data, which is difficult to accurately predict the uncertainty of new energy power stations. SUMMARY

[0004] The present application provides a distributed new energy power station load prediction method to solve the problem of difficulty in accurately predicting the load change and uncertainty of new energy power stations.

[0005] The present application provides a distributed new energy power station load prediction method, which comprises:

[0006] Obtaining load prediction data of new energy power stations in a preset period, the load prediction data comprising historical load data, weather data and power market data;

[0007] Pretreating the load prediction data and constructing a training data set and a test data set suitable for a deep learning model;

[0008] Training the training data set according to a long short-term memory network, a convolutional neural network and a gated recurrent unit respectively to generate a preliminary load prediction model; the preliminary load prediction model comprises a load prediction long short-term memory network model, a load prediction gated recurrent unit model and a load prediction convolutional neural network model;

[0009] Evaluating the performance of the preliminary load prediction model through the test data set to generate a preliminary load evaluation result for judging whether the performance requirement is met;

[0010] If the performance requirement is met, input the prediction result of the preliminary load prediction model into an artificial neural network for weighted average to obtain a stacked ensemble load prediction model;

[0011] Evaluating the performance of the stacked ensemble load prediction model on a validation set to generate a load evaluation result for judging whether the performance requirement is met;

[0012] If the performance requirement is met, the performance of the stacked integrated load prediction model is corrected according to the weather forecast data to output a new energy power station load prediction model for distributed new energy power station load prediction.

[0013] The method based on the deep learning model and the weather forecast data correction model can greatly improve the accuracy and reliability of new energy power station load prediction by intelligently predicting the new energy power station load, and solve the problems of difficult accurate prediction of new energy power station load change and uncertainty.

[0014] Optionally, the historical load data includes historical load data of wind power stations and photovoltaic power stations; the meteorological data includes wind speed, wind direction, air pressure, temperature, humidity of the wind power stations, and sunshine intensity, sun angle, temperature, humidity of the photovoltaic power stations; and the power market data includes electricity price and power demand.

[0015] Optionally, the training method of the weather forecast data correction model includes:

[0016] The historical weather forecast data and meteorological observation data of the new energy power station in a preset period are obtained;

[0017] The historical weather forecast data and the meteorological observation data are preprocessed and divided into weather training data set and weather test data set;

[0018] The weather training data set is trained according to the gated recurrent neural network to generate a preliminary weather forecast data correction model;

[0019] The performance of the preliminary weather forecast data correction model is evaluated by the weather test data set to generate a preliminary weather evaluation result for judging whether the performance requirement is met;

[0020] If the performance requirement is met, the model parameters and hyperparameters are adjusted according to the preliminary weather evaluation result to optimize the training of the weather forecast data correction model.

[0021] The weather forecast data correction model optimized and trained by the above method can receive new weather forecast data in real time, correct the meteorological data input to the new energy power station load prediction model, and improve the accuracy of load prediction.

[0022] Optionally, the load prediction data is preprocessed, and the training data set and test data set suitable for the deep learning model are constructed, including:

[0023] The meteorological data in the load prediction data is extracted as a feature;

[0024] The load prediction data is resampled and corrected according to the feature;

[0025] The outliers in the corrected load forecast data are calculated using the Z-score algorithm, and the outliers are then processed.

[0026] The processed load forecast data is normalized using the Min-Max standardization method to generate a normalized dataset.

[0027] Based on the normalized dataset, construct training and testing datasets suitable for deep learning models.

[0028] Optionally, the algorithm formula for the load prediction long short-term memory network model is as follows:

[0029] f t =σ(W xf ·x t +W hf ·h t-1 +b f );

[0030] i t =σ(W xi ·x t +W hi ·h t-1 +b i );

[0031] o t =σ(W xo ·x t +W ho ·h t-1 +b o );

[0032]

[0033]

[0034] h t =o t ×tanh(c t );

[0035] In the formula, f t i t o t , c t x t h t W represents the vector values ​​of the forget gate, input gate, output gate, candidate memory unit, memory unit, input, and hidden state at time t. xf W xi W xo W xc h represents the weight matrix at time t that parameterizes the forget gate, input gate, output gate, and candidate memory units.t-1 , c t-1 denotes the vector value of the hidden state at time t-1, memory cell, W hf , W hi , W ho , W hc denotes the weight matrix parameterized by the forget gate, the input gate, the output gate, and the candidate memory cell from the hidden state at time t-1 to time t, b f , b i , b o , b c denotes the bias vector of the forget gate, the input gate, the output gate, and the candidate memory cell, σ denotes the sigmoid activation function, tanh denotes the tanh activation function, the symbol “·” is the matrix multiplication operation, and the symbol “×” is the element-wise multiplication operation.

[0036] Optionally, the load prediction gating recurrent unit model algorithm formula is:

[0037] r t = σ(W xr · x t + W hr · h t-1 + b r ) ;

[0038] z t = σ(W xz · x t + W hz · h t-1 + b z ) ;

[0039]

[0040]

[0041] In the formula, r t , z t , h t , x t denote the vector values of the reset gate, the update gate, the candidate hidden state, the hidden state, and the input at time t, W xr , W xz , W xh denote the weight matrix parameterized by the input to the reset gate, the update gate, and the candidate hidden state at time t, h t-1 denotes the vector value of the hidden state at time t-1, W hr , W hz denotes the weight matrix parameterized by the hidden state at time t-1 to the reset gate and the update gate at time t, and W hhb represents the weight matrix for parameterizing the candidate hidden state at time t, obtained by element-wise multiplication of the reset gate at time t and the hidden state at time t-1. r b z b h The bias vectors represent the reset gate, update gate, and candidate hidden states. σ represents the sigmoid activation function, tanh represents the tanh activation function, the symbol "·" represents matrix multiplication, and the symbol "×" represents element-wise multiplication.

[0042] Optionally, the performance of the preliminary load forecasting model is evaluated using the test dataset to generate preliminary load assessment results for determining whether performance requirements are met. This includes: selecting root mean square error and coefficient of determination to evaluate the preliminary load forecasting model, with the evaluation calculation formula as follows:

[0043]

[0044]

[0045] In the formula, RMSE represents the root mean square error, and R0 represents the root mean square error. 2 Let y represent the coefficient of determination, m represent the sample size, and yi represent the true value. Indicates the predicted value. This represents the average predicted value;

[0046] The smaller the root mean square error and the larger the coefficient of determination, the better the performance of the preliminary load prediction model.

[0047] Optionally, the algorithm formula for the artificial neural network is as follows:

[0048] y = Wx + b;

[0049] In the formula, x is the input vector, W is the weight matrix, and b is the bias vector.

[0050] As can be seen from the above technical solutions, this application provides a method for load forecasting of distributed renewable energy power plants. The method includes: acquiring load forecasting data for a preset period of renewable energy power plants, wherein the load forecasting data includes historical load data, meteorological data, and electricity market data; preprocessing the load forecasting data and constructing a training dataset and a test dataset suitable for a deep learning model; training the training dataset using a long short-term memory network, a convolutional neural network, and a gated recurrent unit, respectively, to generate a preliminary load forecasting model; the preliminary load forecasting model includes a load forecasting long short-term memory network model, a load forecasting gated recurrent unit model, and a load forecasting convolutional neural network model; and passing the test dataset... The performance of the preliminary load forecasting model is evaluated on a validation set to generate a preliminary load assessment result to determine whether the performance requirements are met. If the performance requirements are met, the prediction results of the preliminary load forecasting model are input into an artificial neural network for weighted averaging to obtain a stacked integrated load forecasting model. The performance of the stacked integrated load forecasting model is evaluated on a validation set to generate a load assessment result to determine whether the performance requirements are met. If the performance requirements are met, the model is corrected and the performance of the stacked integrated load forecasting model is optimized based on weather forecast data to output a new energy power plant load forecasting model for distributed new energy power plant load forecasting, in order to solve the problem of difficulty in accurately predicting the load changes and uncertainties of new energy power plants. Attached Figure Description

[0051] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the distributed renewable energy power station load forecasting method described in the embodiments of this application;

[0053] Figure 2 This is a schematic diagram of the process for preprocessing load forecasting data in the distributed renewable energy power station load forecasting method described in the embodiments of this application;

[0054] Figure 3 This is a flowchart illustrating the training method for the weather forecast data correction model in the distributed new energy power station load forecasting method described in this application embodiment. Detailed Implementation

[0055] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0056] Distributed renewable energy power stations refer to integrated energy systems distributed on the user side. The construction of distributed renewable energy power stations can alleviate energy supply shortages and reduce environmental pollution.

[0057] In some embodiments, the load forecasting method for distributed renewable energy power stations is based on historical data and empirical judgment, which makes it difficult to accurately predict load changes in renewable energy power stations. On the other hand, the load forecasting method for distributed renewable energy power stations is based on historical load data and meteorological data, but the accuracy of the forecast is affected by a variety of factors, especially the accuracy of weather forecast data, which makes it difficult to accurately predict the uncertainty of renewable energy power stations.

[0058] To address the challenges of accurately predicting load changes and uncertainties at renewable energy power plants, see [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a method for load forecasting of distributed renewable energy power plants. This application provides a method for load forecasting of distributed renewable energy power plants, the method comprising:

[0059] S100: Obtain load forecast data for a preset time period from new energy power plants.

[0060] Load forecasting data includes historical load data, meteorological data, and electricity market data.

[0061] It should be understood that the preset time period can be selected as one year, which not only meets the accuracy of the prediction, but also does not consume too much memory.

[0062] S200: Preprocess the load forecast data and build training and test datasets suitable for deep learning models.

[0063] S300: The training dataset is trained using Long Short-Term Memory Network, Convolutional Neural Network, and Gated Recurrent Unit respectively to generate a preliminary load prediction model.

[0064] The preliminary load forecasting models include a load forecasting long short-term memory network model, a load forecasting gated recurrent unit model, and a load forecasting convolutional neural network model.

[0065] It should be understood that Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network capable of learning long-term dependency information. They address long-term dependency problems by introducing three unique gate structures: an input gate, a forget gate, and an output gate, along with a memory unit. The structure of these gates allows LSTM to selectively remember or forget information, thus effectively managing its short-term and long-term memory. Convolutional Neural Networks (CNNs) are a deep learning architecture, a type of feedforward neural network that includes convolutional computations and has a deep structure. Therefore, they are particularly suitable for processing data with a grid-like structure. Specifically, in this application, historical load data and meteorological data can be used as input sequences, and the current load data can be used as the target sequence. Backpropagation algorithm is used to update the model's weights and biases. During training, optimization algorithms such as gradient descent and Adam optimization can be used to accelerate the training process and improve model performance. The use of LSTM networks, convolutional neural networks, and gated recurrent units ensures the model's prediction accuracy and generalization ability, improving its performance and stability.

[0066] S400: Evaluate the performance of the preliminary load forecasting model using a test dataset to generate preliminary load assessment results to determine whether performance requirements are met.

[0067] Specifically, the performance of the preliminary load prediction model can be evaluated using a validation set, and the hyperparameters and parameters of the model can be adjusted based on the preliminary load assessment results. If the performance requirements are not met, the training dataset can be retrained.

[0068] S500: If the performance requirements are met, the prediction results of the preliminary load forecasting model are input into the artificial neural network for weighted averaging to obtain the stacked integrated load forecasting model.

[0069] It should be understood that Artificial Neural Networks (ANNs) are powerful tools in computer science used to handle machine learning problems. They consist of a set of multilayer perceptrons or neurons, hence the name feedforward neural networks. This network structure is formed by a large number of artificial neurons with hierarchical and interconnected relationships arranged in a specific way. Stacking is an ensemble learning method in machine learning that combines multiple different models to predict and filter results, thereby improving model performance. Since individual deep learning models have low generalization performance, stacking and ensemble artificial neural networks to obtain stacked ensemble load prediction models can automatically adapt weight ratios. The ensemble prediction model not only improves the accuracy of load prediction but also enhances its stability and generalization ability, making it superior to single deep learning models. Furthermore, ensemble training requires less data, enabling short-term model updates.

[0070] S600: Evaluate the performance of the stacked integrated load prediction model on the validation set to generate load assessment results to determine whether performance requirements are met.

[0071] It should be understood that the method for evaluating the performance of the stacked integrated load forecasting model is the same as the method for evaluating the performance of the preliminary load forecasting model, and will not be repeated here.

[0072] S700: If the performance requirements are met, the performance of the stacked integrated load forecasting model is improved and optimized based on weather forecast data to output a new energy power plant load forecasting model for distributed new energy power plant load forecasting.

[0073] It should be understood that applying weather forecast data correction models to load forecasting of renewable energy power plants, by receiving new weather forecast data in real time and correcting the load forecast results, enables real-time monitoring and forecasting of the load of renewable energy power plants, thereby improving the stability and reliability of the power system.

[0074] The above method, based on deep learning models and weather forecast data correction models, significantly improves the accuracy and reliability of load forecasting for new energy power plants by intelligently predicting their loads, thus solving the problem of difficulty in accurately predicting load changes and uncertainties in new energy power plants.

[0075] In some embodiments, the load forecasting model for new energy power plants can be integrated into the monitoring system of new energy power plants to achieve automated and intelligent load forecasting and management.

[0076] It should be understood that integration can be achieved through interface docking or embedded modules.

[0077] In some embodiments, historical load data includes historical load data for wind power plants and photovoltaic power plants; meteorological data includes wind speed, wind direction, air pressure, temperature, and humidity for wind power plants, and solar radiation intensity, solar angle, temperature, and humidity for photovoltaic power plants; and electricity market data includes electricity prices and electricity demand.

[0078] In some embodiments, historical load data, meteorological data, and electricity market data of new energy power plants can be collected using data acquisition devices. These devices can be sensors or data acquisition cards for real-time monitoring and data storage.

[0079] In some embodiments, see Figure 2 , Figure 2 This is a flowchart illustrating the preprocessing of load forecasting data in a distributed renewable energy power plant load forecasting method. The preprocessing of load forecasting data and the construction of training and testing datasets suitable for deep learning models include:

[0080] S210: Extract meteorological data from the load forecast data as features.

[0081] S220: Resample and correct load forecast data based on characteristics.

[0082] In this embodiment, meteorological data can be selected from wind farm power, wind speed, and wind direction; and photovoltaic power station power, temperature, and radiation intensity. The resampling frequency can be one point every 15 minutes, i.e., starting from 0:00, 96 data points per day. High-frequency datasets are filtered by time point, while low-frequency datasets are supplemented using an averaging method.

[0083] S230: Calculate and process outliers in the corrected load forecast data using the Z-score algorithm.

[0084] Specifically, the Z-score algorithm can be used to measure how much a value deviates from the mean in its dataset. The formula is as follows:

[0085]

[0086] In the formula, y is the original data value, μ is the mean of the original data, and σ is the standard deviation of the original data. When the deviation y′ is greater than or equal to 5, it is considered an outlier.

[0087] S240: Normalize the processed load forecast data according to the Min-Max normalization method to generate a normalized dataset.

[0088] It should be understood that data normalization can improve the convergence speed of deep learning models, reduce time costs, and reduce feature misjudgments caused by different feature units and scales.

[0089] Data normalization is performed using the Min-Max standardization method, and the formula is as follows:

[0090]

[0091] In the formula, y is the original data value, min is the minimum value in the dataset, max is the maximum value in the dataset, and y′ is the data after normalization.

[0092] A normalized dataset can be defined as:

[0093]

[0094] In the formula, p i For power data, d i to s i The required meteorological data is given by i, where i is the number of samples. The meteorological data must be available one day earlier than the power data.

[0095] S250: Construct training and testing datasets suitable for deep learning models based on the normalized dataset.

[0096] Based on deep learning algorithms, the values ​​at the same time point on the next day are predicted using data from 7 days at the same time. This dataset is constructed for training and prediction of the deep learning model. The j-th training data point is defined as:

[0097]

[0098] In the formula, j is the starting label.

[0099] In some embodiments, the algorithm formula for the load forecasting long short-term memory network model is:

[0100] f t =σ(W xf ·x t +W hf ·h t-1 +b f );

[0101] i t =σ(W xi ·x t +W hi ·h t-1 +b i );

[0102] o t =σ(W xo ·x t +W ho ·h t-1 +b o );

[0103]

[0104]

[0105] h t =o t ×tanh(c t );

[0106] In the formula, f t i t o t , c t x t h t W represents the vector values ​​of the forget gate, input gate, output gate, candidate memory unit, memory unit, input, and hidden state at time t. xf W xi W xo W xc h represents the weight matrix at time t that parameterizes the forget gate, input gate, output gate, and candidate memory units. t-1 c t-1 W represents the vector value and memory unit of the hidden state at time t-1. hf W hi W ho W hc b represents the weight matrix for parameterizing the hidden state at time t-1 to the forget gate, input gate, output gate, and candidate memory units at time t. f b i b o b c The bias vectors represent the forget gate, input gate, output gate, and candidate memory units. σ represents the sigmoid activation function, tanh represents the tanh activation function, the symbol "·" represents matrix multiplication, and the symbol "×" represents element-wise multiplication.

[0107] In some embodiments, the algorithm formula for the load forecasting gated cyclic unit model is:

[0108] r t =σ(W xr ·x t +W hr ·h t-1 +b r );

[0109] z t =σ(W xz ·x t +W hz ·h t-1 +b z );

[0110]

[0111]

[0112] In the formula, r t z t , h t x t W represents the vector values ​​at time t, including the reset gate, update gate, candidate hidden state, hidden state, and input. xr W xz W xh h represents the weight matrix input to the reset gate, update gate, and candidate hidden state parameterization at time t. t-1 W represents the vector value of the hidden state at time t-1. hr W hz W represents the weight matrix from the hidden state at time t-1 to the reset and update gate parameterization at time t. hh b represents the weight matrix for parameterizing the candidate hidden state at time t, obtained by element-wise multiplication of the reset gate at time t and the hidden state at time t-1. r b z b h The bias vectors represent the reset gate, update gate, and candidate hidden states. σ represents the sigmoid activation function, tanh represents the tanh activation function, the symbol "·" represents matrix multiplication, and the symbol "×" represents element-wise multiplication.

[0113] In some embodiments, the algorithm formula for the load prediction convolutional neural network model is as follows:

[0114] s(i,j)=(n*g)(i,j)=∑ a ∑ b n(i+a,j+b)g(a,b).

[0115] In some embodiments, the algorithm formula for the artificial neural network is:

[0116] y = Wx + b;

[0117] In the formula, x is the input vector, W is the weight matrix, and b is the bias vector.

[0118] In some embodiments, the stacked integrated load prediction model algorithm flow is as follows:

[0119]

[0120] In some embodiments, the performance of the preliminary load forecasting model is evaluated using a test dataset to generate preliminary load assessment results for determining whether performance requirements are met. This includes: selecting root mean square error and coefficient of determination to evaluate the preliminary load forecasting model, with the evaluation calculation formula as follows:

[0121]

[0122]

[0123] In the formula, RMSE represents the root mean square error, and R0 represents the root mean square error. 2 Let y represent the coefficient of determination, m represent the sample size, and yi represent the true value. Indicates the predicted value. This represents the average predicted value;

[0124] The smaller the root mean square error and the larger the coefficient of determination, the better the performance of the preliminary load prediction model.

[0125] In some embodiments, see Figure 3 , Figure 3 This is a flowchart illustrating the training method for a weather forecast data correction model in a distributed renewable energy power plant load forecasting approach. The training method for the weather forecast data correction model includes:

[0126] S710: Obtain historical weather forecast data and meteorological observation data for preset time periods of new energy power plants.

[0127] It should be understood that historical weather forecast data and meteorological observation data for one year can be obtained. The meteorological observation data includes wind speed, wind direction, air pressure, temperature, humidity, solar radiation intensity, and solar angle, among other things.

[0128] S720: Preprocess historical weather forecast data and meteorological observation data, and divide them into weather training datasets and weather test datasets.

[0129] It should be understood that historical weather forecast data and meteorological observation data can be cleaned, organized, and standardized to eliminate noise and outliers and unify the data format and units of different data sources.

[0130] Specifically, the preprocessing can be divided into four steps:

[0131] Step 1, Feature Selection and Resampling: For wind farms, meteorological data such as wind speed and direction are used for correction; for photovoltaic power plants, meteorological data such as temperature and radiation intensity are used for correction. The data sampling frequency is one point every 15 minutes, starting from 0:00, with 96 data points per day. High-frequency datasets are filtered by time point, and low-frequency datasets are supplemented using the averaging method.

[0132] Step 2, Outlier Handling: The Z-score algorithm is used to measure how much a value deviates from the mean in its dataset. The formula is as follows:

[0133]

[0134] In the formula, y is the original data value, μ is the mean of the original data, and σ is the standard deviation of the original data. When the deviation y′ is greater than or equal to 5, it is considered an outlier.

[0135] Step 3, Data Normalization: Normalizing data can improve the convergence speed of deep learning models and reduce time costs. It also reduces feature misclassification caused by differences in feature units and scales. Min-Max normalization is used for data normalization, and the formula is as follows:

[0136]

[0137] In the formula, y is the original data value, min is the minimum value in the dataset, max is the maximum value in the dataset, and y′ is the data after normalization.

[0138] Step 4, Dataset Construction: Through the feature selection, resampling, outlier handling, and data normalization described above, the required dataset is obtained, defined as follows:

[0139]

[0140] In the formula, dr i to sr i For real meteorological data, d i to s i The data represents the predicted meteorological data, and i represents the number of samples.

[0141] By using deep learning algorithms and weather forecast data to refine the model and approximate real-world data, a dataset needs to be constructed for training and prediction of the deep learning model. The j-th training data is defined as:

[0142]

[0143] In the formula, j is the starting label.

[0144] The dataset is divided into a weather training dataset and a weather test dataset for subsequent model training and testing.

[0145] S730: Train the weather training dataset using a gated recurrent neural network to generate a preliminary weather forecast data correction model.

[0146] It should be understood that the algorithm formula for the preliminary weather forecast data correction model is the same as that for the load forecast gated cyclic unit model. The preliminary weather forecast data correction model can learn the relationship between historical weather forecast data and actual load data, and can predict the corrected meteorological model data based on new weather forecast data.

[0147] S740: Evaluate the performance of the preliminary weather forecast data correction model using a weather test dataset to generate preliminary weather assessment results to determine whether the performance requirements are met.

[0148] It should be understood that the method for evaluating the performance of the preliminary weather forecast data correction model is the same as that for evaluating the performance of the preliminary load forecast model, and will not be repeated here. Specifically, the performance of the preliminary weather forecast data correction model can be tested and evaluated using a test set, and the model's hyperparameters and parameters can be adjusted based on the preliminary weather assessment results to improve the model's prediction accuracy and generalization ability. If the performance requirements are not met, the model should be retrained on the weather training dataset.

[0149] S750: If the performance requirements are met, adjust the model parameters and hyperparameters based on the preliminary weather assessment results to optimize the training of the weather forecast data correction model.

[0150] The weather forecast data correction model optimized and trained using the above methods can receive new weather forecast data in real time and correct the meteorological data input to the new energy power plant load forecasting model to improve the accuracy of load forecasting.

[0151] The aforementioned distributed renewable energy power station load forecasting method can be applied to meteorological-related renewable energy power stations, such as wind power stations and photovoltaic power stations. Weather forecast data correction models can be used to improve weather forecasting systems. By using deep learning technology, the dynamic changes of atmospheric systems can be better understood and predicted, thus providing more accurate and timely weather forecasts. This is of great significance for industries such as public safety, agriculture, energy management, and aerospace. Furthermore, this model can also be used for climate change research. By using predictive models and correction algorithms, changes in the climate system can be better simulated and predicted, leading to a better understanding and response to the impacts of climate change. Simultaneously, this model can also be used for early warning and response to natural disasters. By monitoring meteorological data in real time, weather systems that may trigger natural disasters can be identified in a timely manner, allowing necessary measures to be taken in advance to reduce the impact of disasters.

[0152] As can be seen from the above technical solutions, the embodiments of this application provide a method for load forecasting of distributed new energy power plants. The method includes: acquiring load forecasting data of the new energy power plant for a preset period, the load forecasting data including historical load data, meteorological data, and electricity market data; preprocessing the load forecasting data and constructing a training dataset and a test dataset suitable for a deep learning model; training the training dataset according to a long short-term memory network, a convolutional neural network, and a gated recurrent unit to generate a preliminary load forecasting model; the preliminary load forecasting model includes a load forecasting long short-term memory network model, a load forecasting gated recurrent unit model, and a load forecasting convolutional neural network model; evaluating the performance of the preliminary load forecasting model through a test dataset to generate a preliminary load assessment result for determining whether the performance requirements are met; if the performance requirements are met, inputting the prediction result of the preliminary load forecasting model into an artificial neural network for weighted averaging to obtain a stacked integrated load forecasting model; evaluating the performance of the stacked integrated load forecasting model on a validation set to generate a load assessment result for determining whether the performance requirements are met; if the performance requirements are met, correcting the model and optimizing the performance of the stacked integrated load forecasting model according to weather forecast data to output a new energy power plant load forecasting model for distributed new energy power plant load forecasting, thereby solving the problem of difficulty in accurately predicting load changes and uncertainties of new energy power plants.

[0153] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A method for load forecasting of distributed renewable energy power plants, characterized in that, The method includes: Obtain load forecast data for a preset period of time from new energy power plants. The load forecast data includes historical load data, meteorological data, and electricity market data. The load prediction data is preprocessed, and training and testing datasets suitable for deep learning models are constructed. The training dataset is trained using a Long Short-Term Memory Network (LSTM), a Convolutional Neural Network (CNN), and a Gated Recurrent Unit (GRU) to generate a preliminary load prediction model. The preliminary load prediction model includes a load prediction LSTM model, a load prediction GRU model, and a load prediction convolutional neural network model. The performance of the preliminary load forecasting model is evaluated using the test dataset to generate preliminary load assessment results to determine whether the performance requirements are met. If the performance requirements are met, the prediction results of the preliminary load prediction model are input into an artificial neural network for weighted averaging to obtain a stacked integrated load prediction model. The performance of the stacked integrated load prediction model is evaluated on a validation set to generate load assessment results to determine whether the performance requirements are met. If the performance requirements are met, the performance of the stacked integrated load forecasting model is optimized based on weather forecast data to output a new energy power plant load forecasting model for distributed new energy power plant load forecasting.

2. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The historical load data includes historical load data for wind power stations and photovoltaic power stations; the meteorological data includes wind speed, wind direction, air pressure, temperature, and humidity for wind power stations, and solar radiation intensity, solar angle, temperature, and humidity for photovoltaic power stations; the electricity market data includes electricity prices and electricity demand.

3. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The training method for the weather forecast data correction model includes: Acquire historical weather forecast data and meteorological observation data for preset time periods of new energy power plants; The historical weather forecast data and the meteorological observation data are preprocessed and divided into a weather training dataset and a weather test dataset. The weather training dataset is trained using a gated recurrent neural network to generate a preliminary weather forecast data correction model. The performance of the preliminary weather forecast data correction model is evaluated using the weather test dataset to generate preliminary weather assessment results to determine whether the performance requirements are met. If the performance requirements are met, the model parameters and hyperparameters are adjusted based on the preliminary weather assessment results to optimize the training of the weather forecast data correction model.

4. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The load prediction data is preprocessed, and training and testing datasets suitable for deep learning models are constructed, including: Extract the meteorological data from the load forecast data as features; The load forecast data is resampled and corrected based on the described features; The outliers in the corrected load forecast data are calculated using the Z-score algorithm, and the outliers are then processed. The processed load forecast data is normalized using the Min-Max standardization method to generate a normalized dataset. Based on the normalized dataset, construct training and testing datasets suitable for deep learning models.

5. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The algorithm formula for the load prediction long short-term memory network model is as follows: f t =σ(W xf ·x t +W hf ·h t-1 +b f ); i t =σ(W xi ·x t +W hi ·h t-1 +b i ); o t =σ(W xo ·x t +W ho ·h t-1 +b o ); h t =o t ×tanh(c t ); In the formula, f t i t o t , c t x t h t W represents the vector values ​​of the forget gate, input gate, output gate, candidate memory unit, memory unit, input, and hidden state at time t. xf W xi W xo W xc h represents the weight matrix at time t that parameterizes the forget gate, input gate, output gate, and candidate memory units. t-1 c t-1 W represents the vector value and memory unit of the hidden state at time t-1. hf W hi W ho W hc b represents the weight matrix for parameterizing the hidden state at time t-1 to the forget gate, input gate, output gate, and candidate memory units at time t. f b i b o b c The bias vectors represent the forget gate, input gate, output gate, and candidate memory units. σ represents the sigmoid activation function, tanh represents the tanh activation function, the symbol "·" represents matrix multiplication, and the symbol "×" represents element-wise multiplication.

6. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The algorithm formula for the load prediction gated cyclic unit model is as follows: r t =σ(W xr ·x t +W hr ·h t-1 +b r ); z t =σ(W xz ·x t +W hz ·h t-1 +b z ); In the formula, r t z t , h t x t W represents the vector values ​​at time t, including the reset gate, update gate, candidate hidden state, hidden state, and input. xr W xz W xh h represents the weight matrix input to the reset gate, update gate, and candidate hidden state parameterization at time t. t-1 W represents the vector value of the hidden state at time t-1. hr W hz W represents the weight matrix from the hidden state at time t-1 to the reset and update gate parameterization at time t. hh b represents the weight matrix for parameterizing the candidate hidden state at time t, obtained by element-wise multiplication of the reset gate at time t and the hidden state at time t-1. r b z b h The bias vectors represent the reset gate, update gate, and candidate hidden states. σ represents the sigmoid activation function, tanh represents the tanh activation function, the symbol "·" represents matrix multiplication, and the symbol "×" represents element-wise multiplication.

7. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The performance of the preliminary load forecasting model is evaluated using the test dataset to generate preliminary load assessment results to determine whether the performance requirements are met. This includes: selecting root mean square error and coefficient of determination to evaluate the preliminary load forecasting model; the evaluation calculation formula is as follows: In the formula, RMSE represents the root mean square error, and R0 represents the root mean square error. 2 The coefficient of determination is represented by y, where m represents the sample size. i Represents the actual value. Indicates the predicted value. This represents the average predicted value.

8. The distributed renewable energy power station load forecasting method according to claim 1, characterized in that, The algorithm formula for the artificial neural network is as follows: y = Wx + b; In the formula, x is the input vector, W is the weight matrix, and b is the bias vector.

Images