New energy power station power prediction method based on deep learning model

Abstract

The invention discloses a new energy power station power prediction method based on a deep learning model, and the method specifically comprises the following steps: 1, collecting the historical data of a new energy power station, and carrying out the preprocessing of the historical data; step 2, constructing an EMD-CNN-LSTM short-term load prediction model; 3, dividing the historical data processed in the step 1 into a training set and a verification set, and inputting the training set into the EMD-CNN-LSTM short-term load prediction model constructed in the step 2 for training to obtain a trained EMD-CNN-LSTM short-term load prediction model; and step 4, inputting the verification set into the trained EMD-CNN-LSTM short-term load prediction model to obtain a final prediction result. The method solves the problem that an existing prediction method is poor in prediction result accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of new energy power station power prediction, and particularly relates to a new energy power station power prediction method based on a deep learning model. Background Art

[0002] Photovoltaic power generation is a technology that directly converts light energy into electrical energy by using the photovoltaic effect at the semiconductor interface. In recent years, it has been vigorously developed and has reached a certain scale. However, with the large-capacity integration of photovoltaic power generation into the power grid, the power generation power of photovoltaic power stations will inevitably affect the safe and stable operation of the power grid, etc. Therefore, it is necessary to predict the power generation power. The existing prediction methods mainly adopt statistical methods and machine learning methods. Among them, although the statistical method is simple to model, the prediction effect is often poor in accuracy. Another machine learning method, including support vector machines, artificial neural networks, etc., lacks the consideration of the correlation of time series, and the prediction results are also poor in accuracy. Summary of the Invention

[0003] The purpose of the invention is to provide a new energy power station power prediction method based on a deep learning model to solve the problem of poor accuracy of the prediction results of existing prediction methods.

[0004] The technical solution adopted by the invention is that a new energy power station power prediction method based on a deep learning model is specifically implemented according to the following steps:

[0005] Step 1, collect the historical data of the new energy power station and preprocess the historical data;

[0006] Step 2, construct an EMD-CNN-LSTM short-term load prediction model;

[0007] Step 3, divide the historical data processed in Step 1 into a training set and a verification set, input the training set into the EMD-CNN-LSTM short-term load prediction model constructed in Step 2 for training, and obtain a trained EMD-CNN-LSTM short-term load prediction model;

[0008] Step 4, input the verification set into the trained EMD-CNN-LSTM short-term load prediction model to obtain the final prediction result.

[0009] The characteristics of the invention also lie in that

[0010] In Step 1, the historical data includes: total solar radiation O i , air temperature T i , internal temperature IT of photovoltaic modules i , humidity H i , wind speed W i and power generation power P i .

[0011] The process of historical data preprocessing is as follows:

[0012] Step 1.1, form the historical data into a one-dimensional matrix Y i =[O i ,T i ,IT i ,H i ,W i ,P i ;

[0013] Step 1.2, clean the historical data in Step 1.1, and use the Isolation Forest algorithm to denoise the generated power P i , and then normalize the denoised generated power.

[0014] In Step 1.2, the normalization expression is:

[0015]

[0016] In the formula, P i ′ is the denoised generated power, mean is the mean of the power sequence, std is the standard deviation of the power sequence, and d i is the normalized data.

[0017] In Step 2, the EMD-CNN-LSTM short-term load forecasting model consists of an input layer, a CNN layer, an LSTM layer, and an output layer in sequence;

[0018] The input layer decomposes the generated power data processed in Step 1 into multiple groups of modal components through EMD, and then combines the total solar radiation O i , temperature T i , the internal temperature IT of the photovoltaic module i , humidity H i , wind speed W i with each group of modal components, that is, Z i =[O i ,T i ,IT i ,H i ,W i ,d i , as the input of the CNN layer;

[0019] The CNN layer consists of a one-dimensional convolutional layer, a max pooling layer, a one-dimensional convolutional layer, a max pooling layer, and a Dropout layer in sequence. After passing through the CNN layer, a real number vector is obtained as the input of the LSTM layer;

[0020] The information storage unit of the LSTM consists of an input gate, a forget gate, and an output gate;

[0021] The output layer is used to reconstruct the load prediction values of all sequence components output by the LSTM layer and output the prediction results.

[0022] The specific process of decomposing the historical load data processed in Step 1 into multiple groups of modal components through EMD is as follows:

[0023] Step A1: Take the normalized power generation data d(t) as the original ash signal, find its maximum and minimum points, and fit these extreme points to obtain the upper and lower envelopes d min (t) and d max (t);

[0024] Step A2: Calculate the mean of the upper and lower envelopes:

[0025] Step A3: For non-stationary signals, the signal is not monotonically increasing in some regions but has inflection points. If these inflection points are not selected, the two conditions of IMF are not met. Therefore, further screening is required, and Step A4 is executed;

[0026] Step A4: Process the remaining signal through Steps A1 to A3, and stop the processing when SD is less than the threshold value to obtain the first-order modal component IMF1;

[0027] Among them, the expression of the screening threshold value SD is as follows:

[0028]

[0029] Step A5: Subtract the first-order modal component IMF1 from the original ash signal d(t) to obtain the residual term r 1 (t), and then use this residual term r 1 (t) to replace the original ash signal d(t) and repeat the operations from Step A1 to A5. After repeating n times, the required residual value r n (t) is obtained. In this way, after decomposition by EMD, the original ash signal can be expressed as:

[0030]

[0031] The one-dimensional convolutional layer includes a convolutional operation and an activation function. Specifically: the input data is Z i =[O i ,T i ,IT i ,H i ,W i ,d i , where d i is the normalized power generation, and Z i is the corresponding external solar data. Its convolution is defined as:

[0032]

[0033] Wherein, k represents the number of convolution kernels, and W k and b k respectively represent the weights and biases of each convolution kernel, n and m respectively represent the length and number of input data, and g represents the size of the convolution kernel;

[0034] Then, the feature map u = [u 1 ; u 2 ; …; u k obtained by the convolution operation is input into the rectified linear unit to achieve non-linear transformation and obtain the feature map extracted by the convolution layer;

[0035]

[0036] The max pooling layer uses an adaptive pooling method to process the features, and the calculation method of the adaptive pooling is as shown;

[0037] Q = pool(Z, p, s) (9)

[0038] Wherein, Q represents the temporal features output by the CNN network, pool() represents the adaptive pooling function, and p and s respectively represent the size and stride of the pooling; the calculation method of p is obtained by dividing the length of the feature Z by the length of Q, so that each feature Z has the same length after passing through the pooling layer.

[0039] The beneficial effects of the present invention are as follows: The new energy power station power prediction method based on the deep learning model of the present invention predicts the power generation by constructing an EMD-CNN-LSTM short-term load prediction model, fully considering the influence of time series on the power generation of photovoltaic modules, such as seasonality and periodicity, greatly improving the prediction accuracy and reliability. And according to the prediction results, the photovoltaic system can effectively plan the energy distribution, thereby improving the performance and efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the EMD-CNN-LSTM short-term load prediction model in the new energy power station power prediction method based on the deep learning model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0042] Embodiment 1

[0043] The new energy power station power prediction method based on the deep learning model of the present invention is specifically implemented according to the following steps:

[0044] Step 1: Collect the historical data of the new energy power station and preprocess the historical data;

[0045] The historical data includes: total solar radiation O i , air temperature T i , internal temperature of photovoltaic modules IT i , humidity H i , wind speed W i and power generation P i ;

[0046] The process of preprocessing the historical data is as follows:

[0047] Step 1.1: Compose the historical data into a one-dimensional matrix Y i =[O i , T i , IT i , H i , W i , P i ;

[0048] Step 1.2: Clean, denoise, and normalize the historical data in Step 1.1;

[0049] Data cleaning is the process of reviewing and validating a dataset, aiming to delete duplicate information, correct existing errors, and handle invalid values and missing values, etc. By performing data cleaning, the consistency of the data can be maintained, and potential outliers can be detected and processed to ensure that these outliers do not affect model training;

[0050] The following is the specific method of using the Isolation Forest algorithm to denoise the power generation matrix:

[0051] Step S1: Randomly select τ sample points from the power generation dataset X = {P 1 , …, P n} to form a subset X' of X and put it into the root node,

[0052] Step S2: Randomly specify a dimension q from d dimensions, and randomly generate a cut point p in the subset X', min(x ij , j = q, x ij ∈ X') < p < max(x ij , j = q, x ij ∈ X');

[0053] Step S3: This cut point p generates a hyperplane that divides the subset X' space into two subspaces: the sample points with the specified dimension less than p are put into the left child node, and the sample points greater than or equal to p are put into the right child node;

[0054] Step S4, recursively perform Step S2 and Step S3 until all leaf nodes have only one sample point or the isolation tree iTree has reached the specified height;

[0055] Step S5, loop Steps S1 and S4 until n isolation trees iTree are generated;

[0056] Step S6, for each data point x i , let it traverse each isolation tree iTree, calculate the average height of the point in the forest, and then calculate the outlier score of each sample point. The formula for the outlier score is as follows:

[0057]

[0058] Among them,

[0059]

[0060] H(c) = ln(c) + β, β = 0.5772156649 (3)

[0061] In the formula, h(x) represents the number of divisions, that is, the depth, from the root node of the tree to the leaf node where the sample point x is located; E(x) refers to the expected value of each sample point; α is a threshold parameter, c(α) is a constant used to scale the path length; H(c) represents the expected path length when a sample set contains c samples; s(x, α) is the outlier index of x recorded in the isolation tree iTree composed of n samples of training data; the value range of s(x, α) is [0, 1]; the judgment of abnormal situations is as follows;

[0062] The closer s(x, α) is to 1, the higher the possibility that it is an outlier;

[0063] The closer s(x, α) is to 0, the higher the possibility that it is a normal point;

[0064] If s(x, α) of most samples is close to 0.5, it means that there are no obvious outliers in the entire data set and no denoising is required.

[0065] Delete the groups with a high possibility of outliers to obtain the denoised power generation data P i ′;

[0066] Normalize the denoised power generation power matrix, and the expression is as follows:

[0067]

[0068] In the formula, P i' is the denoised power generation, mean is the mean of the power sequence, std is the standard deviation of the power sequence, and d i is the normalized power generation data;

[0069] Step 2, as Figure 1 shown, construct an EMD-CNN-LSTM short-term load forecasting model;

[0070] The EMD-CNN-LSTM short-term load forecasting model consists of an input layer, a CNN layer, an LSTM layer, and an output layer in sequence;

[0071] The input layer decomposes the power generation data processed in Step 1 into multiple groups of modal components through EMD (Empirical Mode Decomposition) to reduce the impact of sequence fluctuations on power generation prediction. Then, the total solar radiation O i , temperature T i , internal temperature of the photovoltaic module IT i , humidity H i , wind speed W i are combined with each group of modal components, that is, Z i = [O i , T i , IT i , H i , W i , d i , as the input of the CNN layer;

[0072] The specific process of decomposing the historical load data processed in Step 1 into multiple groups of modal components through EMD (Empirical Mode Decomposition) is as follows:

[0073] Step A1, take the normalized data d(t) as the original ash signal, find its maximum and minimum points, and fit these extreme points to obtain the upper and lower envelopes d min (t) and d max (t);

[0074] Step A2, find the mean of the upper and lower envelopes:

[0075] Step A3, for non-stationary signals, the signal is not monotonically increasing in some regions but has inflection points. These inflection points can reflect the specific characteristics of the normalized power generation data d(t) (original ash signal). If these inflection points are not selected, the obtained first-order modal function d 1 (t) will be inaccurate, that is, usually it does not meet the two conditions of IMF. Therefore, it is necessary to continue screening and execute Step A4;

[0076] Step A4, for the remaining signal d 1(t) Perform the processing of steps A1 to A3, and stop the processing when SD is less than the threshold value (generally 0.2 - 0.3), so as to obtain the first-order modal component IMF1;

[0077] Among them, the expression of the screening threshold value SD is as follows:

[0078]

[0079] Step A5, subtract the first-order modal component IMF1 from the original ash signal d(t) to obtain the residual term r 1 (t), and then use this residual term r 1 (t) to replace the original ash signal d(t) and repeat the operations of steps A1 to A5. After repeating n times, the required residual value r n (t) is obtained. In this way, after EMD decomposition, the original ash signal can be expressed as:

[0080]

[0081] The CNN layer is successively composed of a one-dimensional convolutional layer, a max pooling layer, a one-dimensional convolutional layer, a max pooling layer, and a Dropout layer;

[0082] The convolutional layer is one of the key components of the convolutional neural network (CNN). It includes two main operations, namely convolutional operation and activation function. Specifically: the input data is Z i =[O i ,T i ,IT i ,H i ,W i ,d i , where d i is the normalized power generation, and Z i is the corresponding external solar data. Its convolution is defined as:

[0083]

[0084] In the formula, k represents the number of convolutional kernels, W k and b k represent the weights and biases of each convolutional kernel respectively, n and m represent the length and quantity of the input data respectively, and g represents the size of the convolutional kernel;

[0085] Then input the feature map u = [u 1 ; u 2 ; …; u k obtained by the convolutional operation into the rectified linear unit to achieve non-linear transformation and obtain the feature map extracted by the convolutional layer;

[0086]

[0087] When the max pooling layer processes the feature Z, it filters out the invalid information through a scaling mapping. Since the feature sizes extracted by convolution kernels of different sizes are different, in order to ensure the consistency of the outputs of multi-scale convolution kernels, the model adopts an adaptive pooling method to process the feature Z, and the calculation method of the adaptive pooling is as shown;

[0088] Q = pool(Z, p, s) (9)

[0089] In the formula, Q represents the temporal feature output by the CNN network, pool() represents the adaptive pooling function, and p and s represent the size and stride of the pooling respectively; the calculation method of p is obtained by dividing the length of the feature Z by the length of Q, so that each feature Z has the same length after passing through the pooling layer;

[0090] In the CNN layer, through a one-dimensional convolutional pooling operation, the input matrix is convolved with a specific convolution kernel to analyze the correlation between climate sequences such as temperature, humidity, and rainfall and the load sequence. In addition, a Dropout layer is also used to avoid overfitting effects;

[0091] The real-valued vector output by the CNN layer will be used as the input of the LSTM layer;

[0092] The LSTM layer uses the autocorrelation function ACF and the partial autocorrelation function PACF to analyze the autocorrelation between historical values and current values, and uses the memory of the LSTM to extract the temporal features of the sequence;

[0093] The information storage unit of the LSTM consists of an input gate, a forget gate, and an output gate to achieve selective storage and extraction of important information in the sequence; these gating units determine the update of the cell state c t and the calculation of the hidden state h t (the output of the LSTM);

[0094] (1) The forget gate determines how much information needs to be discarded from the cell state at the current time step, and the specific calculation is as follows:

[0095] F t = σ(W f * [h t-1 , x t + b f ) (10)

[0096] In the formula, F t is the input of the forget gate, W f is the weight matrix of the forget gate, bf is the bias vector of the forget gate, σ is the sigmoid function, and x t is the input sequence at the current time step, that is, the real-valued vector of the output of the CNN layer;

[0097] Then the cell state c t is updated as follows:

[0098] c t = F t * c t-1 (11)

[0099] (2) The input gate determines how much new information needs to be added to the cell state at the current time step. The specific calculation is as follows:

[0100] I t = σ(W i * [h t-1 , x t + b i ) (12)

[0101] In the formula, I t is the input of the input gate, W i is the weight matrix of the input gate, and b i is the bias vector of the input gate;

[0102] Then the update of the cell state c t is represented as follows:

[0103]

[0104] Among them, the candidate update of the cell state

[0105]

[0106] In the formula, W c is the weight matrix of the candidate update, b c is the bias vector of the candidate update, and tanh is the hyperbolic tangent function;

[0107] (3) The output gate determines how much information of the cell state needs to be output at the current time step. The specific calculation is as follows:

[0108] O t = σ(W o * [h t-1 , x t + b o ) (15)

[0109] In the formula, O t is the input of the output gate, W o is the weight matrix of the input gate, bo is the bias vector of the output gate;

[0110] Then the hidden state h t is updated as follows:

[0111] h t = O t * tanh(c t ) (16)

[0112] The cell state c t is updated under the control of the input gate and the forget gate. The input gate determines the impact of new input information on the cell state, while the forget gate determines the degree to which the previous cell state is retained at the current time step;

[0113] In the output layer, the load prediction values of all sequence components are reconstructed to output the prediction result;

[0114] Step 3: Divide the historical data processed in Step 1 into a training set and a validation set using the cross-validation method. Input the training set into the EMD-CNN-LSTM short-term load prediction model constructed in Step 2 for training. During the training process, obtain the optimal hyperparameters through grid search to get the trained EMD-CNN-LSTM short-term load prediction model;

[0115] Step 4: Input the validation set into the trained EMD-CNN-LSTM short-term load prediction model to obtain the final prediction result.

[0116] Embodiment 2

[0117] In Step 3, during the training process, generate real-time predicted power generation data as shown in Table 1;

[0118] Table 1 Power Generation Data

[0119] Layer(type) Output Shape Param# Conv1d(Conv1D) (None, 1, 32) 15392 Max_pooling1d(MaxPooling1D) (None, 1, 32) 0 Conv1d_1(Conv1D) (None, 1, 64) 10304 Max_pooling1d_1(MaxPooling1D) (None, 1, 64) 0 Conv1d_2(Conv1D) (None, 1, 128) 41088 Max_pooling1d_2(MaxPooling1D) (None, 1, 128) 0 Dropout(Dropout) (None, 1, 128) 0 lstm(LSTM) (None, 1, 200) 263200 lstm_1(LSTM) (None, 1, 100) 120400 Flatten(Flatten) (None, 1, 100) 0 lstm_2(Dense) (None, 1, 100) 80400 dense(Dense) (None, 1, 150) 15150 dense_1(Dense) (None, 1, 1) 604

[0120] It can be obtained from Table 1 that Total param: 546538; Trainable params: 546538; Non-trainable params: 0.

[0121] Embodiment 3

[0122] Model training requires tuning multiple hyperparameters of the network model. Based on the predicted power generation data obtained in Embodiment 2, the hyperparameters will be optimized through experiments below. The hyperparameter optimization table is shown in Table 2.

[0123] Table 2 Hyperparameter Optimization Table

[0124]

[0125] The present invention uses the GridSearchCV in the sklearn data analysis module to implement grid search and cross-validation. The optimal hyperparameters obtained are: Conv1D: 32 64; Hidden layer: 60; Epoch: 300; Batch-size: 64; Optimizer: Adam.

[0126] Example 4

[0127] The training set is used to train the MD-CNN-LSTM short-term load forecasting model, and the validation set is used to validate the trained EMD-CNN-LSTM short-term load forecasting model. Validation is carried out by analyzing the prediction results for error evaluation, where the error evaluation analysis uses the mean absolute error (MAE) and the root mean square error (RMSE).

[0128]

[0129]

[0130] In the formula, N is the number of samples; P c is the true value of the power generation; P c′ is the predicted value of the power generation.

Claims

1. A new energy power station power prediction method based on a deep learning model, characterized in that, it is specifically implemented according to the following steps: Step 1, collect the historical data of the new energy power station and preprocess the historical data; Step 2, construct an EMD-CNN-LSTM short-term load prediction model; Step 3, divide the historical data processed in Step 1 into a training set and a validation set, input the training set into the EMD-CNN-LSTM short-term load prediction model constructed in Step 2 for training, and obtain a trained EMD-CNN-LSTM short-term load prediction model; Step 4, input the validation set into the trained EMD-CNN-LSTM short-term load prediction model to obtain the final prediction result.

2. The new energy power station power prediction method based on a deep learning model according to claim 1, characterized in that, In Step 1, the historical data includes: total solar radiation O i , air temperature T i , internal temperature of the photovoltaic module IT i , humidity H i , wind speed W i and power generation P i .

3. The new energy power station power prediction method based on a deep learning model according to claim 2, characterized in that, the process of historical data preprocessing is: Step 1.1, form the historical data into a one-dimensional matrix Y i = [O i , T i , IT i , H i , W i , P i ; Step 1.2: Clean the historical data in Step 1.1 and perform normalization on the generated power P i Use the Isolation Forest algorithm for denoising, and then normalize the generated power after denoising.

4. The new energy power station power prediction method based on a deep learning model according to claim 3, characterized in that, in Step 1.2, the normalization expression is: Where P i ′ is the denoised power generation power, mean is the mean of the power sequence, std is the standard deviation of the power sequence, and d i is the normalized data.

5. The new energy power station power prediction method based on a deep learning model according to claim 2, characterized in that, in Step 2, the EMD-CNN-LSTM short-term load prediction model is successively composed of an input layer, a CNN layer, an LSTM layer and an output layer; The input layer decomposes the power generation data processed in step 1 into multiple groups of modal components through EMD, and then combines the total solar radiation O i , air temperature T i , the internal temperature IT of the photovoltaic module i , humidity H i , wind speed W i with each group of modal components, that is, Z i = [O i , T i , IT i , H i , W i , d i , which is used as the input of the CNN layer; The CNN layer is successively composed of a one-dimensional convolutional layer, a max pooling layer, a one-dimensional convolutional layer, a max pooling layer, and a Dropout layer. After passing through the CNN layer, a real number vector is obtained as the input of the LSTM layer; The information storage unit of the LSTM is composed of an input gate, a forget gate and an output gate; The output layer is used to reconstruct the load prediction values of all sequence components output by the LSTM layer and output the prediction result.

6. The new energy power station power prediction method based on a deep learning model according to claim 5, characterized in that, the specific process of decomposing the historical load data processed in Step 1 into multiple groups of modal components by EMD is as follows: Step A1: Take the normalized power generation data d(t) as the original ash signal, find its maximum and minimum points, and fit these extreme points to obtain the upper and lower envelopes d min (t) and d max (t); Step A2, calculate the mean value of the upper and lower envelopes: Step A3, for non-stationary signals, the signal is not monotonically increasing in some regions, but inflection points appear. If these inflection points are not selected, the two conditions of IMF are not satisfied, so screening needs to be continued and Step A4 is executed; Step A4, process the remaining signal through Steps A1 to A3, and stop processing when SD is less than the threshold value, so as to obtain the first-order modal component IMF1; Among them, the expression of the screening threshold value SD is as follows: Step A5: Subtract the first-order modal component IMF1 from the original ash signal d(t) to obtain the residual term r 1 (t), and then use this residual term r 1 (t) to replace the original ash signal d(t) and repeat the operations of steps A1 to A5. After repeating n times, the required residual value r n (t) is obtained. In this way, after EMD decomposition, the original ash signal can be expressed as:

7. The new energy power station power prediction method based on a deep learning model according to claim 5, characterized in that, The one-dimensional convolutional layer includes a convolutional operation and an activation function, specifically: the input data is Z i =[O i , T i , IT i , H i , W i , d i , where d i is the normalized power generation, and Z i is the corresponding external solar data, and its convolution is defined as: where k represents the number of convolutional kernels, W k and b k represent the weights and biases of each convolutional kernel respectively, n and m represent the length and number of input data respectively, and g represents the size of the convolutional kernel; Then, the feature map u = [u 1 ; u 2 ;...; u k obtained from the convolution operation is input into the rectified linear unit function to achieve a non-linear transformation and obtain the feature map extracted by the convolutional layer; the max pooling layer uses an adaptive pooling method to process features, and the calculation method of adaptive pooling is shown; Q = pool(Z, p, s) (9) Wherein, Q represents the temporal features output by the CNN network, pool() represents the adaptive pooling function, and p and s respectively represent the size and stride of pooling; the calculation method of p is obtained by dividing the length of the feature Z by the length of Q, so that each feature Z has the same length after passing through the pooling layer.

Images