A Hybrid Deep Learning Method for Short-Term Irradiance Prediction
By combining data preprocessing, signal decomposition, error compensation framework and encoding codec model, the nonlinearity and non-stationarity problem of the difficulty in predicting irradiance of a single deep learning model is solved, and higher prediction accuracy and effective processing of long sequence input is achieved.
Patent Information
- Application Number
- CN202211027097.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-25
- Publication Date
- 2025-07-01
- Estimated Expiration
- 2042-08-25
AI Technical Summary
A single deep learning prediction model is difficult to accurately predict nonlinear and non-stationary changes in irradiation, and existing models based on RNN and CNN are difficult to take into account long-sequence input and long-term dependence, resulting in a decrease in prediction accuracy.
The hybrid deep learning short-term irradiance prediction method is adopted to improve the prediction effect through the combination of data preprocessing, signal decomposition, error compensation framework and encoding codec model. Specific steps include obtaining historical irradiance and meteorological data, data preprocessing, irradiance sequence decomposition, updating the encoded codec model based on the error compensation framework, and making predictions.
It effectively reduces the nonlinearity and non-stationarity of irradiance prediction and improves prediction accuracy, especially in terms of long-sequence input and long-term dependence, which significantly improves the prediction performance of the model.
Smart Images
Figure CN115409258B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a hybrid deep learning short-term irradiance prediction method, belonging to the technical field of photovoltaic power generation. Background Art
[0002] Solar energy resources are the most promising renewable energy sources. A survey by the International Renewable Energy Agency shows that as of 2020, 29% of global electricity production comes from renewable energy sources, among which solar energy accounts for 26.77% of the renewable energy power generation and is increasing year by year. However, due to the uncertainty and intermittency of irradiance, photovoltaic power generation shows considerable instability, which increases the difficulty of grid connection and scheduling of photovoltaic power generation and restricts the wide application of solar energy resources.
[0003] Currently, there are already various irradiance prediction methods based on deep learning, such as using models like LSTM and CNN for irradiance prediction. Some scholars are dedicated to the research on hybrid prediction methods of RNN models and CNN models, such as the LSTM-CNN combined prediction model. However, the CNN model is better at extracting spatial features and has limited ability to extract time-dependent features; the RNN model can maintain temporal dependence but is difficult to process long input sequences. The above characteristics indicate that deep learning models based on RNN and CNN are difficult to balance long sequence input and long-term dependence, and the model needs to be improved. In addition, due to the non-linearity and non-stationarity of irradiance, a single deep learning model is difficult to accurately predict the change of irradiance. There is still room for improvement in the deep learning prediction method by fusing models at the data end and the error end. Summary of the Invention
[0004] Technical Problem:
[0005] The technical problem to be solved by the present invention is that due to the non-linearity and non-stationarity of irradiance, a single deep learning prediction model is difficult to achieve a satisfactory prediction effect; in addition, the existing deep learning models based on RNN and CNN are difficult to balance long sequence input and long-term dependence, resulting in the prediction model being difficult to make full use of historical information for irradiance prediction, making it difficult to guarantee the model accuracy when the prediction step size becomes larger; finally, the machine learning model predicts irradiance as an approximation of the true distribution, and it is inevitable to generate errors in the approximation process of the machine learning model, and these errors are difficult to eliminate under this machine learning model, resulting in components that can still be predicted in the prediction results.
[0006] To solve the above technical problems, the present invention provides a hybrid deep learning short-term irradiance prediction method, which fuses models from three perspectives: data processing, model optimization, and error processing. The prediction effect is improved through multi-model fusion.
[0007] Technical Solution:
[0008] The present invention provides a hybrid deep learning short-term irradiance prediction method, which includes the following steps:
[0009] S1. Training data acquisition: Obtain historical irradiance data and its corresponding meteorological data in the target area, and make a supervised data set according to the prediction task;
[0010] S2. Data preprocessing, including meteorological information feature encoding and data normalization, specifically including the following:
[0011] S3. Irradiance sequence decomposition: Use the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm to decompose the irradiance sequence into several subsequences, so as to reduce the non-stationarity and non-linearity of the decomposed subsequences;
[0012] S4. Update the encoder-decoder model based on the error compensation framework;
[0013] S5. Prediction: Input the historical data into the error compensation framework in step S4 and the trained machine learning model to predict the solar irradiance in the future for multiple steps.
[0014] Further, the step S1 includes the following:
[0015] (1.1) Obtain historical irradiance data and its corresponding meteorological data (including but not limited to temperature, humidity, air pressure, wind speed, etc.) in the target area;
[0016] (1.2) If a certain segment of the historical data is missing or illegal, use the mean value of the adjacent front and rear data to replace it to ensure the continuity and authenticity of the data, so as to ensure the quality of the training data;
[0017] (1.3) Match the supervised information to make a supervised data set. By reading the irradiance from the current time period t0 to t N as the supervised information, match the historical irradiance with the corresponding meteorological information as the input information to make a supervised data set.
[0018] Further, the step S2 specifically includes the following:
[0019] (2.1) Encode the meteorological information corresponding to the irradiance. Use the one-hot encoding method to encode the weather type, and use the numerical value itself as the encoding value for numerical information;
[0020] (2.2) To ensure reasonable model gradient changes during training, it is necessary to perform a normalization operation on the input data. The normalization formula is as follows:
[0021]
[0022] Among them, a represents the feature in the dataset, a' represents the normalized feature value, and a max and a min respectively represent the maximum and minimum values of this feature in the historical data.
[0023] Furthermore, the detailed steps of the signal decomposition algorithm in step S3 are as follows:
[0024] (3.1) Add different Gaussian white noises to the input irradiance sequence I(t) to obtain multiple noisy sequences: I i (t) = I(t) + ε·w i (t), i = 1, …, K, where ε is the standard deviation of the noise, and w i (t) is different white noise, and K is the number of added noises;
[0025] (3.2) The first decomposition mode IMF1 can be expressed as: where represents the empirical mode decomposition value of the sequence with the i-th noise added, and E j (·) represents the j-th component generated by the empirical mode decomposition operator;
[0026] (3.3) Calculate the residual r k (t) = r k-1 (t) - IMF k (t); where the initial condition r0(t) = I(t);
[0027] (3.4) The remaining modes
[0028] Furthermore, the characteristics of the error compensation framework in step S4 are as follows:
[0029] (4.1) This framework is an end-to-end solar irradiance prediction framework; after inputting historical irradiance and historical meteorological data, this framework can automatically run and output multi-step prediction results, and at the same time, it can automatically utilize error information inside the framework without other operations;
[0030] (4.2) In the error acquisition stage, this framework only uses the encoder-decoder model to obtain error information, and the model parameters are not updated during this process; in the error compensation stage, the error information is used to dynamically update the model;
[0031] (4.3) Without changing the structure of the encoder-decoder model, this framework can effectively reduce the irradiance prediction error.
[0032] Furthermore, the detailed steps of the error compensation framework are as follows:
[0033] (1) Obtain subsequences using a sliding window:
[0034] This framework first receives the data from time T-2K to T-1 as input, and starting from time T-2K, it is divided into K subsequences in the form of a sliding window with a sliding step of 1;
[0035] The K subsequences obtained are respectively {(I T-2k , …, I T-K-1 ), …, (I t-K+1 , …, I t ), …, (I T-K-1 , …, I T-2 )}, where I t represents the solar irradiance at time t;
[0036] (2) Rolling to obtain error information:
[0037] For the subsequence (I t-K , …, I t-1 ), its corresponding external meteorological information is (M t-K , …, M t-1 ). The input to the encoder-decoder model under the current parameters is {(I t-K , …, I t-1 ), (M t-K , …, M t-1 ), and its predicted value is Its corresponding supervision information can be expressed as (I t , …, I t+N ), where N represents the prediction step; According to the predicted value and the supervision information, the prediction error E t of the model at the current moment can be obtained = [e t , …, e t+N ;
[0038] Loop through the above K subsequences and obtain the prediction error of the encoder-decoder model under the current parameters at the current moment through the supervision information; This loop process can obtain an error sequence (E T-K , …, E T-1 );
[0039] Moreover, the model parameters are not updated during this process;
[0040] (3) Information fusion:
[0041] In this stage, the error sequence (E T-K , …, E T-1 ) is fused with the historical irradiance and meteorological data to obtain a new feature input {(E T-K , …, E T-1 ), (I T-K , …, I T-1 ), (M T-K , …, MT-1 )}, and its supervision information is (I T , …, I T+N );
[0042] The input features after information fusion add the prediction error information of the current model at the corresponding historical moment. This information is determined by the internal prediction mechanism of the model and is fed back to the encoder-decoder model.
[0043] (4) Update the model:
[0044] Update the model parameters using the input features and supervision information in step (3).
[0045] Furthermore, the structure of the encoder-decoder in step S4 is as follows:
[0046] Train the encoder-decoder model. Use the irradiance read in the current time period t0 to t N as the supervision information, and the historical irradiance and meteorological information before t0 as the input data to train the encoder-decoder model; furthermore, the encoder-decoder in S4 includes the following:
[0047] The encoder consists of a cascaded structure of a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The TCN is responsible for obtaining long sequence inputs and maintaining temporal dependencies. The compressed short sequence passes through the LSTM to maintain temporal dependencies. The encoder structure first receives long sequence inputs through several layers of TCN (the number of layers depends on the length of the input sequence). Secondly, the feature sequence extracted by the TCN is compressed and output to the LSTM. Finally, the output of the LSTM is used as the encoded output of the encoder.
[0048] The decoder consists of a series structure of a long short-term memory network LSTM and a multi-layer perceptron MLP. The series length is determined by the prediction step; and the performance of multi-step outputs is balanced through a loss function. The decoder first receives the output of the encoder. The LSTM is responsible for decoding, and the decoded output is sent to the MLP, where the MLP is used to match the output dimension.
[0049] Furthermore, the characteristics of the encoder-decoder are as follows:
[0050] 1) The encoder is composed of a cascade of TCN and LSTM; the TCN receives long sequence inputs, extracts features through multiple layers to obtain an output sequence, and intercepts the latter segment of the sequence as the input of the LSTM;
[0051] 2) The decoder is composed of a cascade of LSTM and MLP; the LSTM receives the encoder state, and the MLP is used for dimension matching of the LSTM output;
[0052] 3) Each layer of the TCN is composed of TCN residual blocks, and each TCN residual block contains a concatenation module and a residual connection; the concatenation module is composed of two groups of identical dilated causal convolutional layers, weight normalization layers, ReLu activation units, and dropout layers;
[0053] 4) By designing a loss function, the prediction performance of multi-step outputs is balanced, and the loss function is designed as follows:
[0054]
[0055] where K represents the number of prediction steps, loss i represents the output loss of the i-th step, α i represents the weight of loss i , w represents the parameters of the model, and β represents the regularization coefficient.
[0056] Furthermore, the TCN and LSTM in the encoder-decoder include:
[0057] 1. The temporal convolution operation can be expressed by the following formula:
[0058]
[0059] where * represents the convolution operation, d represents the dilation coefficient, χ represents the input sequence, s represents an element of the sequence, f represents the convolution kernel, k represents the size of the convolution kernel, and s - d·i represents the element selected by the dilated convolution;
[0060] 2. TCN residual block:
[0061] o1 = dropout(ReLU(Norm(F(s)))),
[0062] o2 = dropout(ReLU(Norm(O1))),
[0063] O tcn = s + O2,
[0064] where Norm represents weight normalization, ReLU represents the activation function, and dropout represents the dropout layer; O tcn represents the output of the TCN residual block, that is, the output of each layer of the TCN;
[0065] 3. The LSTM neural network model includes:
[0066] Forget gate f t :
[0067] f t = sigmoid(W if x t + b if + Whf h t-1 + b hf ),
[0068] Input gate i t :
[0069] i t = sigmoid(W ii x t + b ii + W hi h t-1 + b hi ),
[0070] Activation function g t :
[0071] g t = tanh(W ii x t + b ii + W hi h t-1 + b hi ),
[0072] Output gate o t :
[0073] o t = sigmoid(W io x t + b io + W ho h t-1 + b ho ),
[0074] The state c of the memory cell corresponding to the current moment t :
[0075]
[0076] The output state h of the LSTM cell t :
[0077]
[0078] Where W if and b if respectively represent the weight matrix and bias matrix of the external input of the forget gate, W hf and b hf respectively represent the weight matrix and bias matrix of the hidden state input of the forget gate, W ii and b ii respectively represent the weight matrix and bias matrix of the external input of the input gate, W hi and b hiThey respectively represent the weight matrix and bias matrix for the input of the input gate hidden state, \(W\) io and \(b\) io They respectively represent the weight matrix and bias matrix for the external input of the output gate, \(W\) ho and \(b\) ho They respectively represent the weight matrix and bias matrix for the input of the output gate hidden state, \(f\) t 、\(i\) t and \(o\) t They are respectively the outputs of the forget gate, input gate, and output gate at time \(t\), \(h\) t is the hidden state at time \(t\).
[0079] Beneficial effects:
[0080] The signal decomposition part in the present invention can effectively reduce the non - linearity and non - stationarity of the irradiance sequence, thereby effectively enhancing the predictable components in the irradiance sequence.
[0081] In the encoder part of the present invention, taking advantage of the characteristic of TCN having a long receptive field to obtain long - sequence inputs, after the output of TCN, short - sequence inputs are intercepted and input into LSTM, enabling LSTM to obtain longer historical information in the case of short - sequence inputs; the encoder has the ability to handle long - sequence inputs and maintain the temporal dependence of time features, and is more suitable for feature extraction of time series; the decoder structure can ensure that the model sequentially outputs multiple prediction steps, and balances the prediction performance of multi - step outputs through the loss function. On the premise of only relying on historical information, the encoder - decoder structure of the present invention can achieve high - precision irradiance prediction results.
[0082] The error compensation framework of the present invention can effectively reduce the irradiance prediction error by using the prediction error of the model at the current parameters for historical adjacent moments as additional input features without changing the structure of the machine - learning model; moreover, the error compensation framework in the present invention is a general paradigm, which can be used for different types of machine - learning prediction models, and can effectively utilize error information for different machine - learning prediction models to dynamically update the model according to the error information; in addition, this framework can simultaneously complete irradiance prediction and error compensation through only an end - to - end structure. Description of the Drawings
[0083] Figure 1 is a schematic diagram of the irradiance prediction process of the present invention;
[0084] Figure 2 is a flowchart of the signal decomposition algorithm of the present invention;
[0085] Figure 3 is a structural diagram of the encoder - decoder of the present invention;
[0086] Figure 4Schematic diagram of the internal TCN of the codec of the present invention;
[0087] Figure 5 Schematic diagram of the internal LSTM of the codec of the present invention. Detailed implementation manners
[0088] In order to more clearly illustrate the technical solutions in the present invention, the present invention will be described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.
[0089] Embodiment 1
[0090] Reference Figure 1 , a deep learning multi-step irradiance prediction method based on a codec, comprising the following steps:
[0091] S1, training data acquisition, acquiring historical irradiance data of a target area and its corresponding meteorological data and making a supervised data set. Further, S1 includes the following contents:
[0092] (1.1) Acquiring historical irradiance data of a target area and its corresponding meteorological data. In this embodiment, temperature T, humidity H, air pressure P, wind speed W, etc. are selected as meteorological data;
[0093] (1.2) If a certain segment of historical data is missing or illegal, the average value of adjacent front and rear data is used to replace it to ensure the continuity and authenticity of the data and to guarantee the quality of the training data;
[0094] (1.3) Matching supervision information to make a supervised data set. By reading the irradiance x0 to x N at the current time period t0 to t N as the supervision information for multi-step prediction, matching the historical irradiance with the corresponding meteorological information as the input information, and making a supervised data set; for example: selecting the first 24 moments before the current moment as the historical information to input into the model and predicting the irradiance in the next 6 moments, then the data set can be expressed as ([x -24 , T -24 , W -24 , P -24 , H -24 , …, x -1 , T -1 , W -1 , P -1 , H -1 ; [x0, …, x5]), where [x0, …, x5] is the supervision information, and [x -t , T -t , W -t , P -t , H -tIt represents the irradiance and meteorological information at t moments before the current moment.
[0095] S2. Data preprocessing, including meteorological information feature encoding and data normalization, specifically including the following:
[0096] (2.1) Encode the meteorological information corresponding to the irradiance. Use the one - hot encoding method to encode the weather type, and use the numerical value itself as the encoding value for numerical information.
[0097] (2.2) To ensure reasonable model gradient changes during training, it is necessary to perform a normalization operation on the input data. The normalization formula is as follows:
[0098]
[0099] Where a represents the feature in the dataset, a' represents the normalized feature value, a max and a min respectively represent the maximum and minimum values of this feature in the historical data; in this embodiment, the features that need to be normalized are historical irradiance information and historical meteorological information, where the historical meteorological information includes temperature, humidity, wind speed, and air pressure.
[0100] S3. Irradiance sequence decomposition:
[0101] Use the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm to decompose the irradiance sequence into several subsequences, so as to reduce the non - stationarity and non - linearity of the decomposed subsequences.
[0102] S4. Update the encoder - decoder model based on the error compensation framework. The schematic diagram of the error compensation framework is shown in the appendix Figure 1 , and its content includes:
[0103] (4.1) Obtain the irradiance prediction error of the encoder - decoder model corresponding to the current parameters from time T - K to T - 1. The model is not updated during this process; the prediction error at each moment is obtained by using the data of the previous K moments before this moment as the model input.
[0104] (4.2) Perform information fusion on the prediction error sequence from time T - K to T - 1 and the external meteorological data sequence, and use it to update the model.
[0105] S5. Prediction. Input the historical data into the error compensation framework in step S4 and the trained encoder - decoder model to predict the future multi - step solar irradiance.
[0106] Furthermore, the flowchart of the signal decomposition algorithm in step S3 is shown in the appendix Figure 2 , and its algorithm steps are as follows:
[0107] (3.1) The input irradiance sequence I(t) is added with different Gaussian white noises to obtain multiple noisy sequences: I i (t) = I(t) + ε·w i (t), i = 1, …, K, where ε is the standard deviation of the noise, and w i (t) is different white noise, and K is the number of different noises added;
[0108] (3.2) The first decomposed mode IMF1 can be expressed as: where represents the empirical mode decomposition value of the sequence with the i-th noise added, and E j (·) represents the j-th component generated by the empirical mode decomposition operator;
[0109] (3.3) Calculate the residual r k (t) = r k-1 (t) - IMF k (t); where the initial condition r0(t) = I(t);
[0110] (3.4) The remaining modes
[0111] Furthermore, the characteristics of the error compensation framework in step S4 are as follows:
[0112] (4.1) This framework is an end-to-end solar irradiance prediction framework; after inputting historical irradiance and historical meteorological data, this framework can automatically run and output prediction results with multiple time steps, and at the same time, it can automatically utilize error information inside the framework without other operations;
[0113] (4.2) In the error acquisition stage, this framework only uses the encoder-decoder model to obtain error information, and the model parameters are not updated during this process; in the error compensation stage, the error information is used to dynamically update the model;
[0114] (4.3) Without changing the structure of the encoder-decoder model, this framework can effectively reduce the irradiance prediction error.
[0115] Furthermore, in this embodiment, the method for making the dataset in step S1 is as follows:
[0116] In this embodiment, the historical irradiance information and historical meteorological data information at the previous 24 moments of moment t are selected as the input of the model, and the irradiance at the next 6 moments is predicted. Then, the input data of the prediction model dataset can be expressed in the following form, {(I t-24 , …, I t-1 ), (T t-24 , …, T t-1 ), (W t-24 , …, Wt-1 ),(P t-24 ,…,P t-1 ),(H t-24 ,…,H t-1 )}, the supervision information of the dataset can be expressed as (I0, …, I5), where [I t-m , T t-m , W t-m , P t-m , H t-m represents the irradiance and meteorological information at the previous m moments at time t;
[0117] In addition, to meet the working mechanism of the error compensation framework, error information needs to be added to the dataset; for the current moment T, its error information comes from the error compensation framework of S4; at other moments, it represents the error acquisition stage, and there is a lack of error information during this process. To ensure the consistency of the model input form, this embodiment introduces an error sequence of zeros; the actual effect is equivalent to ignoring the compensation function of the encoder-decoder model during the error acquisition process.
[0118] Furthermore, the detailed steps of the error compensation framework in step S4 are as follows:
[0119] (1) Obtain subsequences by sliding window:
[0120] This framework first receives the data from T - 48 to T - 1 as input, and starting from the moment T - 48, it is divided into 24 subsequences in the form of a sliding window with a sliding step of 1;
[0121] The 24 subsequences obtained are respectively {(I T-48 , …, I T-25 ), …, (I t-23 , …, I t ), …, (I T-25 , …, I T-2 )}, where I t represents the solar irradiance at time t;
[0122] (2) Obtain error information by rolling:
[0123] For the subsequence (I t-24 , …, I t-1 ), its corresponding external meteorological information is (M t-24 , …, M t-1 ). In this embodiment, the external meteorological information M includes temperature T, humidity H, air pressure P, and wind speed W, and can be expressed as (T t-24 , …, T t-1 ), (W t-24 , …, W t-1 ), (P t-24 , …, Pt-1 ), (H t-24 , …, H t-1 );
[0124] The input of the encoder - decoder model under the current parameters is \(\{(I t-24 , …, I t-1 ), (M t-24 , …, M t-1 )\}, and its output is the predicted value The corresponding supervision information can be expressed as \((I t , …, I t+5 ), where \(N\) represents the prediction step - length; according to the predicted value and the supervision information, the prediction error \(E\) of the model at the current moment can be obtained t = [e t , …, e t+5 ;
[0125] Loop through the above 24 subsequences and obtain the prediction error of the encoder - decoder model at the current moment under the current parameters through the supervision information; this loop process can obtain the error sequence \((E T-24 , …, E T-1 );
[0126] Moreover, the model parameters of the encoder - decoder are not updated during this process;
[0127] (3) Information fusion:
[0128] In this stage, the error sequence \((E T-24 , …, E T-1 ) is fused with the historical irradiance and meteorological data to obtain a new feature input \(\{(E T-24 , …, E T-1 ), (I T-24 , …, I T-1 ), (M T-24 , …, M T-1 )\}, and its supervision information is \((I T , …, I T+5 );
[0129] The input features after information fusion add the prediction error information of the current model at the corresponding historical moments. This information is determined by the internal prediction mechanism of the model and is fed back to the encoder - decoder model;
[0130] (4) Update the model:
[0131] Update the model parameters of the encoder - decoder using the input features and supervision information in step (3).
[0132] Furthermore, the structure of the encoder - decoder model is as follows:
[0133] The encoder consists of a cascaded structure of a temporal convolutional network (TCN) and a long short-term memory network (LSTM). The TC block is responsible for obtaining long sequence inputs and maintaining temporal dependencies. The compressed short sequence maintains temporal dependencies through the LSTM. The encoder structure first receives long sequence inputs through several layers of TCN (the number of layers depends on the length of the input sequence). Secondly, the feature sequence extracted by the TCN is compressed and output to the LSTM (here the LSTM only selects part of the output of the TCN layer as the input). Finally, the output of the LSTM is used as the encoded output of the encoder.
[0134] The decoder consists of a series structure of a long short-term memory network LSTM and a multi-layer perceptron MLP. The series length is determined by the prediction step size; and the performance of multi-step outputs is balanced through a loss function. The decoder first receives the output of the encoder, and the LSTM is responsible for decoding. After decoding, it is output to the MLP, where the MLP is used for dimension matching with the output.
[0135] Furthermore, the encoder-decoder in step S4 is shown in the appendix Figure 3 , and its characteristics are as follows:
[0136] 1) The encoder is composed of a cascade of TCN and LSTM; the TCN receives long sequence inputs, extracts features through multiple layers to obtain an output sequence, and intercepts the latter segment of the sequence as the input of the LSTM;
[0137] 2) The decoder is composed of a cascade of LSTM and MLP; the LSTM receives the encoder state, and the MLP is used for dimension matching of the LSTM output;
[0138] 3) As Figure 4 shown, the TCN residual block consists of two cascaded modules and a residual connection; each cascaded module contains a dilated causal convolutional layer, a weight normalization layer, a ReLu activation unit, and a dropout layer;
[0139] 4) By designing a loss function, the prediction performance of multi-step outputs is balanced, and the loss function is designed as follows:
[0140]
[0141] where K represents the number of prediction step sizes, loss i represents the output loss of the i-th step size, α i represents the weight of loss i , w represents the parameters of the model, and β represents the regularization coefficient; in this embodiment, the outputs at different times are regarded as equally important, so the weight coefficient α i takes the same value.
[0142] Furthermore, the TCN and LSTM structures in step S4 are as Figure 5 , and specifically include:
[0143] 1. The temporal convolution operation can be expressed by the following formula:
[0144]
[0145] where * represents the convolution operation, d represents the dilation coefficient, χ represents the input sequence, s represents an element of the sequence, f represents the convolution kernel, k represents the size of the convolution kernel, and s - d·i represents the element selected by the dilated convolution;
[0146] 2. TCN residual block:
[0147] O1 = dropout(ReLU(Norm(F(s)))),
[0148] O2 = dropout(ReLU(Norm(O1))),
[0149] O tcn = s + O2,
[0150] where Norm represents weight normalization, ReLU represents the activation function, and dropout represents the dropout layer; O tcn represents the output of the TCN residual block, that is, the output of each layer of the TCN;
[0151] 3. The LSTM neural network model includes:
[0152] The forget gate, which is used to discard unimportant information; the output f of the forget gate in the forgetting stage t is calculated by the sigmoid activation function, and whether to discard the information of the previous moment is determined by the output value of the activation function. The calculation formula of the forget gate is as follows:
[0153] f t = sigmoid(W if x t + b if + W hf h t-1 + b hf ),
[0154] The input gate and the activation function jointly complete the memory selection, and their matrix multiplication determines which values are saved to the current state;
[0155] The calculation formula of the input gate i t is as follows:
[0156] i t = sigmoid(W ii x t + b ii + W hi h t-1 + b hi),
[0157] Activation function g t has the following calculation formula:
[0158] g t = tanh(W ii x t + b ii + W hi h t-1 + b hi ),
[0159] The state c of the memory cell corresponding to the current moment t is jointly determined by the input gate and the forget gate. The matrix multiplication of the forget gate and the previous moment's state represents discarding some unnecessary information, and the matrix multiplication of the input gate and the activation function represents preserving important information. Its calculation formula is as follows:
[0160]
[0161] Output gate o t :
[0162] o t = sigmoid(W io x t + b io + W ho h t-1 + b ho ),
[0163] The output gate determines the output value of the memory cell state at the current moment, and thus obtains the output state h t of the LSTM cell. Its calculation formula is as follows:
[0164]
[0165] Among them, W if and b if respectively represent the weight matrix and bias matrix of the external input of the forget gate, W hf and b hf respectively represent the weight matrix and bias matrix of the hidden state input of the forget gate, W ii and b ii respectively represent the weight matrix and bias matrix of the external input of the input gate, W hi and b hi respectively represent the weight matrix and bias matrix of the hidden state input of the input gate, W io and b io respectively represent the weight matrix and bias matrix of the external input of the output gate, W ho and b ho respectively represent the weight matrix and bias matrix of the hidden state input of the output gate, ft , i t and o t are the outputs of the forget gate, input gate, and output gate at time t, respectively, and h t is the hidden state at time t.
Claims
1. A hybrid deep learning short-term irradiance prediction method, characterized in that, It includes the following steps: S1. Training data acquisition: Obtain historical irradiance data of the target area and its corresponding meteorological data, and make a supervised data set according to the prediction task; S2. Data preprocessing, including meteorological information feature encoding and data normalization; S3. Irradiance sequence decomposition: Use the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm to decompose the irradiance sequence into several subsequences, so as to reduce the non-stationarity and non-linearity of the decomposed subsequences; S4. Update the encoder-decoder model based on the error compensation framework; S5. Prediction: Input the historical data into the error compensation framework in step S4 and the trained machine learning model to predict the solar irradiance in the future for multiple steps; The detailed steps of the signal decomposition algorithm in step S3 are as follows: (3.1) Add different Gaussian white noises to the input irradiance sequence I(t) to obtain multiple noisy sequences: I i (t) = I(t) + ε·w i (t), i = 1, …, K, where ε is the standard deviation of the noise, and w i (t) is different white noise, and K is the number of different noises added; (3.2) The first decomposed mode IMF1 is expressed as: where represents the empirical mode decomposition value of the sequence with the i-th noise added, and E j (·) represents the j-th component generated by the empirical mode decomposition operator; (3.3) Calculate the residual r k (t) = r k-1 (t) - IMF k (t); where the initial condition r0(t) = I(t); (3.4) Other modes The characteristics of the error compensation framework in step S4 are as follows: (4.1) This framework is an end-to-end solar irradiance prediction framework; after inputting historical irradiance and historical meteorological data, this framework can automatically run and output prediction results for multiple time steps, and at the same time, it can automatically utilize error information inside the framework without other operations; (4.2) In the error acquisition stage, this framework only uses the encoder-decoder model to obtain error information, and the model parameters are not updated during this process; The error compensation stage uses the error information to dynamically update the model; (4.3) Without changing the structure of the encoder-decoder model, this framework can effectively reduce the irradiance prediction error; The detailed steps of the error compensation framework are as follows: (1) Obtain subsequences using a sliding window: This framework first receives the data from time T - 2K to time T - 1 as input, and starting from time T - 2K, it is divided into K subsequences in the form of a sliding window with a sliding step of 1; Obtain K subsequences as {(I T-2k , …, I T-K-1 ), …, (I t-K+1 , …, I t ), …, (I T-K-1 , …, I T-2 )}, where I t represents the solar irradiance at time t; (2) Obtain error information by rolling: For the subsequence (I t-K , …, I t-1 ), the corresponding external meteorological information is (M t-K , …, M t-1 ). The input of the codec model under the current parameters is {(I t-K , …, I t-1 ), (M t-K , …, M t-1 )}, and its predicted value is The corresponding supervision information is represented as (I t , …, I t+N ), where N represents the prediction step length; the prediction error E t = [e t , …, e t+N of the model at the current moment is obtained according to the predicted value and the supervision information; Loop through the above K subsequences and obtain the prediction error of the encoder-decoder model at the current moment under the current parameters through the supervision information; this loop process results in an error sequence (E T-K ,…,E T-1 ); Moreover, the model parameters are not updated during this process; (3) Information fusion: This stage fuses the error sequence (E T-K , …, E T-1 ) with historical irradiance and meteorological data to obtain a new feature input {(E T-K , …, E T-1 ), (I T-K , …, I T-1 ), (M T-K , …, M T-1 )}, and its supervision information is (I T , …, I T+N ); The input features after information fusion add the prediction error information of the current model for the corresponding historical time. This information is determined by the internal prediction mechanism of the model, and this information is fed back to the encoder-decoder model; (4) Update the model: Use the input features in step (3) and the supervision information to update the model parameters.
2. The prediction method according to claim 1, characterized in that, Step S1 includes the following content: (1.1) Obtain historical irradiance data of the target area and its corresponding meteorological data, including but not limited to temperature, humidity, air pressure, and wind speed; (1.2) If a certain segment of the historical data is missing or illegal, use the mean value of the adjacent front and rear data to replace it to ensure the continuity and authenticity of the data, so as to ensure the quality of the training data; (1.3) Match the supervision information to create a supervision dataset. By reading the irradiance from the current time period t0 to t N as the supervision information, match the historical irradiance and the corresponding meteorological information as the input information, and create a supervision dataset.
3. The prediction method according to claim 1, wherein Step S2 specifically includes the following content: (2.1) Encode the meteorological information corresponding to the irradiance. Use the one-hot encoding method to encode the weather type, and use the numerical value itself as the encoding value for numerical information; (2.2) To ensure reasonable changes in the model gradient during the training process, perform a normalization operation on the input data. The normalization formula is as follows: Among them, a represents the feature in the dataset, a' represents the normalized feature value, a max and a min respectively represent the maximum and minimum values of this feature in the historical data.
4. The prediction method according to claim 1, wherein The structure of the encoder-decoder in step S4 is as follows: Train the encoder-decoder model, using the irradiance read during the current period t0 to t N as the supervision information, and the historical irradiance and meteorological information before the time t0 as the input data to train the encoder-decoder model; further, the encoder-decoder described in S4 includes the following: The encoder consists of a cascaded structure of a Temporal Convolutional Network (TCN) and a Long Short-Term Memory Network (LSTM). The TCN is responsible for obtaining long-sequence inputs and maintaining temporal dependencies. The compressed short sequence is passed through the LSTM to maintain temporal dependencies. First, the encoder structure receives long-sequence inputs through several layers of TCN, where the number of layers depends on the length of the input sequence. Secondly, the feature sequence extracted by the TCN is compressed and output to the LSTM. Finally, the output of the LSTM is used as the encoded output of the encoder. The decoder consists of a concatenated structure of cascaded Long Short-Term Memory Networks (LSTM) and Multi-Layer Perceptrons (MLP). The concatenated length is determined by the prediction step size, and the performance of multi-step outputs is balanced through a loss function. The decoder first receives the output of the encoder. The LSTM is responsible for decoding, and the decoded output is then sent to the MLP, which is used to match the output dimension.
5. The prediction method according to claim 4, wherein The characteristics of the encoder-decoder are as follows: 1) The encoder is composed of a cascade of TCN and LSTM. The TCN receives long-sequence inputs, extracts features through multiple layers to obtain an output sequence, and the latter segment of the sequence is intercepted as the input to the LSTM. 2) The decoder is composed of a cascade of LSTM and MLP. The LSTM receives the encoder state, and the MLP is used for dimension matching of the LSTM output. 3) Each layer of the TCN consists of TCN residual blocks, which contain a concatenation module and a residual connection. The concatenation module consists of two groups of identical dilated causal convolutional layers, weight normalization layers, ReLU activation units, and dropout layers. 4) By designing a loss function, the prediction performance of multi-step outputs is balanced. The loss function is designed as follows: where K represents the number of predicted steps, and loss i represents the output loss at the i-th step, α i represents the weight of loss i , w represents the parameters of the model, and β represents the regularization coefficient.
6. The prediction method according to claim 4, wherein The TCN and LSTM in the encoder-decoder include:
1. The temporal convolution operation is expressed as follows: where * represents the convolution operation, d represents the dilation coefficient, χ represents the input sequence, s represents an element of the sequence, f represents the convolution kernel, k represents the size of the convolution kernel, and s - d·i represents the element selected by the dilated convolution.
2. TCN residual block: o1 = dropout(ReLU(Norm(F(s)))) o2 = dropout(ReLU(Norm(O1))) O tcn = s + O2, where Norm represents weight normalization, ReLU represents the activation function, and dropout represents the dropout layer; O tcn represents the output of the TCN residual block, that is, the output of each layer of the TCN; 3. The LSTM includes: Forgotten gate f t : f t = sigmoid(W if x t + b if + W hf h t-1 + b hf ), Input gate i t : i t = sigmoid(W ii x t + b ii + W hi h t-1 + b hi ), Activation function g t : g t = tanh(W ii x t + b ii + W hi h t-1 + b hi ), Output gate o t : o t = sigmoid(W io x t + b io + W ho h t-1 + b ho ), The state c of the memory cell corresponding to the current moment t : LSTM cell output state h t : Among them, W if and b if represent the weight matrix and bias matrix of the external input of the forget gate respectively. W hf and b hf represent the weight matrix and bias matrix of the hidden state input of the forget gate respectively. W ii and b ii represent the weight matrix and bias matrix of the external input of the input gate respectively. W hi and b hi represent the weight matrix and bias matrix of the hidden state input of the input gate respectively. W io and b io represent the weight matrix and bias matrix of the external input of the output gate respectively. W ho and b ho represent the weight matrix and bias matrix of the hidden state input of the output gate respectively. f t 、i t and o t are the outputs of the forget gate, input gate and output gate at time t respectively. h t is the hidden state at time t.