Solar irradiance prediction method based on multi-scale convolution echo state network

Abstract

The invention provides a solar irradiance prediction method based on a multi-scale convolution echo state network, and the method comprises the steps: 1, obtaining the solar irradiance data of different time scales in different regions according to different set time and places; 2, dividing the solar irradiance data into four parts according to seasons, and then dividing the solar irradiance data into two parts according to training data and test data; step 3, combining the multi-scale convolutional neural network MCNN and the echo state network ESN to construct a multi-scale convolutional echo state network MCESN model; step 4, the MCNN in the MCESN model is trained, and the output connection matrix Wout of the ESN in the MCESN model is solved; and 5, performing solar irradiance prediction by using the MCESN model to obtain a prediction result. According to the method, the MCESN prediction model with high prediction precision and high training speed is designed; and the MCNN is adopted to extract solar irradiance characteristic information on different time scales so as to improve the prediction precision. And encoding and fitting the extracted feature information by adopting the ESN so as to improve the prediction efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of solar irradiance prediction, and in particular to a solar irradiance prediction method based on a multi-scale convolutional echo state network. Background Art

[0002] In recent years, solar irradiance prediction has played an important role in renewable energy systems. Solar energy is a sustainable, environmentally friendly, and more importantly, inexhaustible resource with potential that is unmatched by other energy sources. However, due to the mobility of clouds and the variability of meteorological factors, solar energy is highly nonlinear, non-stationary, and intermittent. Undoubtedly, this uncertainty will have a certain impact on the safe and stable operation of solar energy systems. Therefore, it is of great significance to accurately and efficiently predict solar energy to ensure the safety, efficiency, and reliability of renewable energy systems.

[0003] At present, various solar irradiance prediction technologies have been successfully proposed and patents have been applied for, such as: Publication No. CN117151303A, the invention name is a method and system for ultra-short-term solar irradiance prediction based on a hybrid model; Publication No. CN116451598A, the invention name is a solar irradiance prediction method based on a denoising diffusion probability model; Publication No. CN116167465A, the invention name is a solar irradiance prediction method based on multivariate time series ensemble learning; Publication No. CN110781458A, the invention name is a method based on a hybrid regression model A method for predicting surface solar irradiance; the publication number is CN114201924A, and the invention name is a solar irradiance prediction method and system based on transfer learning; the publication number is CN116245011A, and the invention name is a solar irradiance prediction model based on Transformer-LSTM and error correction; the publication number is CN115689055A, and the invention name is a short-term solar irradiance prediction method and device; the publication number is CN110956270A, and the invention name is a solar irradiance prediction method based on XGBOOST algorithm, etc.

[0004] These patent applications use different models to predict solar irradiance, but the models are mainly concentrated in feedforward neural networks and traditional recursive neural networks. Although feedforward neural networks, especially deep feedforward neural networks, have good learning capabilities, the model itself lacks certain dynamic characteristics and memory performance, and it is difficult to capture the complex long-term dependencies between input signals and target signals. Although traditional recursive neural networks have good dynamic characteristics and memory performance, they need to use the gradient descent algorithm to adjust all connection weights during training, and there are inherent problems in the model, such as low solution rate, heavy computational burden, and easy to fall into local optimality. This, to a certain extent, limits its application in actual solar irradiance prediction, which has high requirements for modeling efficiency. Summary of the invention

[0005] Under comprehensive consideration of modeling accuracy and modeling efficiency, the present invention provides a solar irradiance prediction method based on a multi-scale convolution echo state network (MCESN), which effectively integrates the powerful feature extraction capability of a multi-scale convolutional neural network (MCNN) and the efficient training mechanism of an echo state network (ESN). On the one hand, a multi-scale convolutional neural network is used as a pre-training link to fully mine and extract the implicit feature information inside the solar irradiance to improve the accuracy of solar irradiance prediction. On the other hand, an echo state network is used as a prediction link to replace the fully connected layer of the multi-scale convolutional neural network to encode and fit the above-extracted solar irradiance feature information to enhance the efficiency of solar irradiance prediction.

[0006] A solar irradiance prediction method based on a multi-scale convolutional echo state network comprises the following steps:

[0007] First, a multi-scale convolutional neural network feature extraction pre-training link is established. In order to fully mine and extract the internal feature information of solar irradiance, the present invention adopts an improved convolutional neural network as a pre-training link, namely a multi-scale convolutional neural network MCNN. MCNN introduces a multi-scale convolution kernel, which overcomes the problems of limited modeling ability of a single convolutional neural network, limited local receptive field, and indifferent calculation of input time series information. Therefore, MCNN can mine and extract the implicit feature information inside solar irradiance on different time scales to improve the accuracy of solar irradiance prediction.

[0008] Secondly, an echo state network prediction link is established. In order to improve the modeling efficiency, the present invention adopts an echo state network instead of the fully connected layer of MCNN as the prediction link to encode and fit the above-extracted solar irradiance feature information. ESN is a new type of recursive neural network, which is characterized by replacing the hidden layer with a reserve pool composed of a large number of sparsely connected nodes, and only the output connection weights need to be trained. The reserve pool can map the input signal from the low-dimensional input space to the high-dimensional state space, and use a linear regression algorithm to fit the expected value in the output layer. Therefore, ESN has a simple and efficient training mechanism, which can improve the efficiency of solar irradiance prediction.

[0009] Finally, a multi-scale convolution echo state network prediction model is established and its performance is evaluated. On the basis of the above, the MCESN prediction model is established. This model not only has the powerful feature extraction ability of MCNN but also has the efficient training mechanism of ESN, which can be better used for solar irradiance prediction. When using MCESN to predict solar irradiance, it includes two stages: pre-training stage and prediction stage. In the pre-training stage, by calculating the mean square error between feature information and label value, the size of MCNN convolution kernel is adjusted by stochastic gradient descent method, and then multi-dimensional solar irradiance feature information on different time scales is extracted. In the prediction stage, the extracted feature information is loaded into ESN as input information, and the expected solar irradiance value is finally fitted in the output layer through the mapped state information. In addition, the performance of MCESN is quantitatively analyzed and verified by four indicators: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R).

[0010] The beneficial effects of the present invention are:

[0011] 1. Design a MCESN prediction model with high prediction accuracy and fast training speed; use MCNN to extract solar irradiance feature information on different time scales to improve prediction accuracy. Use ESN to encode and fit the extracted feature information to improve prediction efficiency.

[0012] 2. Compared with traditional methods, the proposed MCESN prediction model can have smaller prediction error and higher prediction efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is the flow chart of solar irradiance prediction method based on MCESN;

[0014] Figure 2 This is the structure diagram of the MCNN model in solar irradiance prediction based on MCESN;

[0015] Figure 3This is the structure diagram of the ESN model in the solar irradiance prediction based on MCESN;

[0016] Figure 4 This is the structure diagram of the solar irradiance prediction model based on MCESN. DETAILED DESCRIPTION

[0017] The present invention is further described below in conjunction with specific embodiments, but the protection scope of the present invention is not limited thereto.

[0018] A solar irradiance prediction method based on multi-scale convolutional echo state network, such as Figure 1 As shown, the following steps are included:

[0019] Step 1: According to the set time and location, obtain the solar irradiance data of different regions and different time scales.

[0020] Step 2: Divide the solar irradiance data into four parts according to season, and then divide it into two parts according to training data and test data.

[0021] Step 3: Combine the multi-scale convolutional neural network (MCNN) with the echo state network (ESN) to construct a multi-scale convolutional echo state network (MCESN) model.

[0022] Step 4: Train the MCNN in the MCESN model and output the connection matrix W for the ESN in the MCESN model out 's seeking.

[0023] Step 5: Use the MCESN model to predict solar irradiance and obtain the prediction results.

[0024] Step 6: Compare the prediction results of MCESN with other models.

[0025] The following is a further explanation of the above steps:

[0026] For step 1, since the solar irradiance is negligible after sunset and before sunrise, and the longitude and latitude of each region are different, the solar irradiance will also be different, so the historical data obtained comes from three different regions from 8 am to 5 pm every day.

[0027] For step 2, due to the influence of seasons, the dataset is divided into four different parts. The first and second months of each season are used as training sets, and the third month is used as a test set.

[0028] For step 3, MCESN includes MCNN and ESN:

[0029] (1) Multi-scale Convolutional Network (MCNN)

[0030] CNN is a multi-layer perceptron network, which originated from the study of cat visual cortex and has the characteristics of weight sharing and local connection. Solar irradiance is a time series, and time series prediction is a one-dimensional data input. Therefore, a one-dimensional CNN can be used to predict solar irradiance. Multi-scale CNN (MCNN) shows good advantages in feature extraction. It is composed of multiple single-scale CNN (SCNN) modules in parallel. Figure 2 The structure diagram of the MCNN model in the solar irradiance prediction based on MCESN. It includes four steps: padding, convolution, padding, and pooling. When the sequence is input, 0 is padded on the left and right sides of the sequence, and then the padded sequence is convolved to obtain the feature information matrix. Then the padding layer is added before the maximum pooling layer. In order to ensure that there is output during the convolution process, a condition is required for padding before MAX pooling: when the sequence length is an even number, 0 is added to the front and back of the sequence respectively, otherwise, 0 is only added to the front of the sequence. Finally, the MAX pooling operation is performed.

[0031] Assume that MCESN has m SCNN modules, each SCNN consists of Figure 1 The single convolution process of i sequences in the present invention is obtained by connecting them in series. In this invention, m = 1, 2, 3, i = 1, 2, 3. For each SCNN, assume that the input signal is The data length is L 1m ,k m Indicates the convolution kernel length of the mth module, n i Represents the number of convolution kernels in the i-th layer. The convolution kernel of the i-th layer of the m-th SCNN is expressed as:

[0032]

[0033] Each row of the matrix represents the weight of each convolution kernel, which is randomly initialized and updated iteratively through back propagation. The formula for the first filling is:

[0034]

[0035] The result after the i-th convolution is:

[0036]

[0037] The first convolution formula is:

[0038]

[0039] where j = 2, 3, ..., L 1m +k1+2, is the output of the i-th convolution layer, and f is the RELU activation function. The characteristic signal obtained after the i-th convolution layer is as follows:

[0040]

[0041] The padding result before the i-th maximum pooling is:

[0042]

[0043] Where g∈N * The formula for the i-th maximum pooling is:

[0044]

[0045] The result of the i-th maximum pooling is:

[0046]

[0047] Where g∈N * For convenience of representation, the signal of the next convolutional layer is redefined as follows:

[0048]

[0049] Where L (t+1)m It is M im The column length of the i-th convolution is:

[0050]

[0051] Note that after the first convolution, the signal is no longer one-dimensional information. Therefore, the i-th convolution formula is rewritten as:

[0052]

[0053] where j = 2, 3, ..., L tm -k i +2.

[0054] Feature information at different time scales can be extracted through three SCNN channels with different convolution kernels, denoted as M 31 、M 32 and M 33 In addition, an effective SCNN should satisfy L 2m -k3+3+1≥2 is a necessary and sufficient condition. The final output Y of MCNN is fitted as follows:

[0055] Y=∑WX′+b⒀

[0056] Where X′={M 31 ,M 32 ,M 33}, b is the bias.

[0057] (2) Echo State Network (ESN)

[0058] Echo State Network (ESN) is a new type of recurrent neural network consisting of an input layer, a hidden recurrent layer (called a dynamic reservoir), and an output layer. In ESN, the input signal, reservoir state, and output signal are represented as u(i), x(i), and y(i), respectively.

[0059] The dynamic reservoir plays a core role in ESN. It contains a large number of sparsely connected neurons whose connection weights are randomly initialized and remain unchanged during training and testing. Therefore, the local minimum avoids the problems caused by gradient descent training to a certain extent. On the one hand, the dynamic reservoir encodes the input signal from the low-dimensional input space to the high-dimensional state space; on the other hand, the dynamic reservoir performs complex nonlinear transformations on the input signal. Then, a supervised learning mechanism is used to train the output connection weights W out ESN only needs to update the connection weights from the reservoir to the output layer. Therefore, compared with the classic recurrent neural network, the training process of ESN is simple and very fast.

[0060] Then, the best feature information extracted above is concatenated and passed to the echo state network as the input signal. The dynamic reservoir further encodes the feature information into a high-dimensional state space. Finally, the target solar irradiance is fitted with the echo state network in the output layer through a linear regression algorithm. Therefore, MCESN can perform better than the classic ESN due to the feature extraction in the pre-training stage. Figure 3 This is the structure diagram of the ESN model in solar irradiance prediction based on MCESN.

[0061] The entire workflow of MCESN includes the following steps:

[0062] 1) First, MCNN is used to mine solar irradiance data at different time scales and extract feature information.

[0063] 2) Then the best feature information extracted above is concatenated and passed to ESN as input signal.

[0064] 3) The dynamic reservoir further encodes the characteristic information into the high-dimensional state space, and finally fits the target solar irradiance through ESN in the output layer.

[0065] Assume that ESN has K, N1, and L neurons in the input layer, hidden layer, and output layer, respectively. They represent the input weight connection matrix, the internal hidden weight connection matrix, and the output weight connection matrix respectively. The reservoir state and network output equations can be expressed as:

[0066] x(i+1)=tanh(W in u(i+1)+W res x(i)) ⒁

[0067] y(i+1)=W out x(i+1) ⒂

[0068] In ESN, the main work is to train the output connection weight matrix. In the learning phase, training is performed according to the teacher's forced signal. At the same time, the updated reserve pool state is collected into the matrix G. Let l r Represents the length of the training data set, then the state matrix of the reserve pool is recorded as:

[0069]

[0070] The corresponding teacher vector matrix can be expressed as:

[0071]

[0072] According to the least squares ridge regression approximate solution, the value of the output matrix can be obtained:

[0073] (Wout) T =(G T G+λE) -1 G T D ⒅

[0074] E is the identity matrix of order N1, and λ represents the regularization parameter, which requires a lot of experiments to determine the appropriate value for a given learning task. After obtaining the output weight matrix, the ESN can be tested.

[0075] For step 4, in the MCNN network, because the weight of the convolution kernel is randomly assigned by normal distribution, in order to obtain reliable feature information, it is necessary to connect the feature information of different time scales of MCNN, calculate the mean square error, set the learning rate to 0.00001, and then let the MCNN network back propagate to learn the weight parameters of the convolution kernel, and then use Adam to optimize and adjust the weight parameters of the convolution kernel. In ESN, the linear relationship between the matrix composed of the data updated by the reserve pool state matrix and the teacher label matrix is ​​used to calculate the output connection matrix Wout using the ridge regression method. The process of calculating Wout is the training process of ESN.

[0076] For step 5, Figure 4 The structure diagram of the solar irradiance prediction model based on MCESN. In the pre-training stage, the input data is divided into nine data as one time scale and three time scales as one input data. Each time scale data is subjected to padding, convolution, and pooling operations. After three rounds of operations, M is obtained. 31 、M 32 and M 33 . Put M 31 、M 32 and M 33Expand the series to get the time series X′ of feature information, where X′ is the feature information obtained by inputting data in the pre-training stage. In the prediction stage, the feature information is used as input, and W in , x(i+1) is calculated by the internal connection relationship between Wres and the size of the feature information. The prediction result can be calculated through the linear relationship between Wout and x(i+1).

[0077] For step 6, in order to detect the accuracy of the MCESN prediction results, four different indicators, namely root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and correlation coefficient (R), are used to evaluate the prediction accuracy of MCESN. These indicators are expressed by the following formulas:

[0078]

[0079] where y predicted , i is the predicted solar irradiance, y real , i is the actual measured solar irradiance, cov(y predicted ,y real ) is the covariance between the predicted value and the actual measured value, and is the standard deviation between the predicted value and the actual measured value.

[0080] Experimental results analysis

[0081] The experimental data are selected from the hourly solar irradiance records of three stations, Blythe Ne, Seeley and Owens Lake South, in the California Irrigation Management Information System (CIMIS). Affected by the season, the experimental data are divided into four seasons: spring, summer, autumn and winter. Table 1 gives the statistical results of the predicted values ​​and measured values ​​of the hourly solar irradiance of the ESN model and the MCESN model at the three stations in four seasons. The results of RMSE, MAE, MAPE and R indicators show that when the reserve pool size is the same, the prediction error of MCESN is smaller than that of ESN, which reflects the generalization ability and effectiveness of MCESN. It can be seen from Table 2 that the performance of the MCESN model proposed in the present invention in the hourly solar irradiance prediction task is better than that of BP, RF, LSTM, DESN and MCNN models.

[0082] Table 1 Statistical comparison of the measured and predicted values ​​of MCESN and ESN models at Blythe Ne, Seeley and Owens Lake South stations in four seasons

[0083]

[0084]

[0085] Table 2 Statistical comparison of the measured and predicted values ​​of MCESN and other models at Blythe Ne station in two seasons

[0086]

[0087]

[0088] The above is only a preferred embodiment of the present invention, which is used to illustrate the technical solution of the present invention, but not to limit the present invention. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present invention, some of the technical features therein are modified and replaced, and they will not depart from the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A solar irradiance prediction method based on a multi-scale convolutional echo state network, characterized in that: The following steps are involved: Step 1: According to the set time and place, obtain the solar irradiance data of different regions and different time scales; Step 2: Divide the solar irradiance data into four parts according to season, and then divide it into two parts according to training data and test data; Step 3: Combine the multi-scale convolutional neural network MCNN and the echo state network ESN to build a multi-scale convolutional echo state network MCESN model; Step 4: Train the MCNN in the MCESN model and output the connection matrix W for the ESN in the MCESN model out of seeking; Step 5: Use the MCESN model to predict solar irradiance and obtain the prediction results.

2. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 1 is characterized in that: In step 1, the solar irradiance data obtained are from three different regions from 8 am to 5 pm every day.

3. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 1 is characterized in that: In step 2, the solar irradiance data are divided into four parts according to season, with the first and second months of each season as training sets and the third month as test sets.

4. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 1 is characterized in that: The multi-scale convolutional neural network MCNN is composed of multiple SCNN modules in parallel. The SCNN module is a single-scale CNN. The MCNN includes four steps: padding, convolution, padding, and pooling. When a sequence is input, 0 is padded on the left and right sides of the sequence, and then a convolution operation is performed on the padded sequence to obtain a feature information matrix. Then, a padding layer is added before the maximum pooling layer. The padding before MAX pooling requires a condition: when the sequence length is an even number, 0 is added to the front and back of the sequence respectively, otherwise, 0 is only added to the front of the sequence, and finally a MAX pooling operation is performed.

5. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 4 is characterized in that: The MCESN model has m SCNN modules, each SCNN is obtained by connecting i sequences of single convolution processes in series. For each SCNN, the input signal is The data length is L 1m , k m Indicates the convolution kernel length of the mth module, n i represents the number of convolution kernels in the i-th layer, and the convolution kernel in the i-th layer of the m-th SCNN is expressed as: Each row of the matrix represents the weight of each convolution kernel. The convolution kernel is randomly initialized and updated iteratively through back propagation. The formula for the first filling is: The result after the i-th convolution is: The first convolution formula is: where j = 2, 3, ..., L 1m +k1+2, is the output of the i-th layer of convolution, and f is the RELU activation function; the characteristic signal obtained after the i-th layer of convolution is as follows: The padding result before the i-th maximum pooling is: Where g∈N * ; The formula for the i-th maximum pooling is: The result of the i-th maximum pooling is: Where g∈N * ; The signal of the next convolutional layer is redefined as follows: Where L (t+1)m It is M im The column length of the i-th convolution is: The i-th convolution formula is rewritten as: where j = 2, 3, ..., L tm -k i +2; The feature information at different time scales is extracted through three SCNN channels with different convolution kernels, denoted as M 31 、M 32 and M 33 ; An effective SCNN should satisfy L 2m -k3+3+1≥2, the final output Y of MCNN is fitted as follows: Y=∑WX′+b⒀ Where X′={M 31 ,M 32 ,M 33 }, b is the bias.

6. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 5 is characterized in that: The echo state network (ESN) is a recursive neural network consisting of an input layer, a hidden loop layer, and an output layer. In the ESN, the input signal, the reservoir state, and the output signal are represented as u(i), x(i), and y(i), respectively. ESN has K, N1 and L neurons in the input layer, hidden layer and output layer respectively. They represent the input weight connection matrix, the internal hidden weight connection matrix and the output weight connection matrix respectively. The reservoir state and network output equations are expressed as: x(i+1)=tanh(W in u(i+1)+W res x(i)) ⒁ y(i+1)=W out x(i+1) ⒂ In the learning phase, training is performed according to the teacher's mandatory signal; at the same time, the updated reserve pool state is collected into the matrix G, assuming l r Represents the length of the training data set, then the state matrix of the reserve pool is recorded as: The corresponding teacher vector matrix is ​​expressed as: According to the least squares ridge regression approximate solution, find the value of the output matrix: (W ou t ) T =(G T G+λE) -1 G T D⒅E is the identity matrix of order N1, and λ represents the regularization parameter.

7. The solar irradiance prediction method based on multi-scale convolutional echo state network according to claim 6 is characterized in that: In step 5, in the pre-training stage, the input data is divided into nine data as one time scale and three time scales as one input data, and each time scale data is subjected to padding, convolution, and pooling operations respectively; after three rounds of operations, M is obtained. 31 、M 32 and M 33 ; put M 31 、M 32 and M 33 Expand the series to get the time series X′ of feature information, where X′ is the feature information obtained by inputting data in the pre-training stage; in the prediction stage, the feature information is used as input, and W in , W res The internal connection relationship with the size of the feature information is used to calculate x(i+1); the prediction result is calculated through the linear relationship between Wout and x(i+1).

Images