Photovoltaic power prediction method based on adaptive correction quantile regression neural network

Abstract

The invention discloses a photovoltaic power prediction method based on an adaptive modified quantile regression neural network. Feature extraction is realized by constructing a double-flow hybrid neural network so as to improve prediction accuracy, a branch, combined with a multi-head attention mechanism, of the convolutional neural network is responsible for extracting long-term features, a branch of a bidirectional gating circulation unit focuses on identifying short-term fluctuation, and the double-flow hybrid neural network is combined with quantile regression. In order to solve the problems of quantile crossing and non-differentiable zero point of a loss function, a self-adaptive correction marble loss function is provided, and smooth function optimization is introduced to ensure monotone increasing of predicted quantiles and whole-domain differentiable of the loss function. According to the method, point prediction, interval prediction and probability density prediction can be realized, the prediction effect is verified through a multi-dimensional evaluation index, potential information of photovoltaic power is fully mined, and the method has practical engineering application value.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction technology, and in particular to a photovoltaic power prediction method based on an adaptive modified quantile regression neural network. Background Technology

[0002] With dwindling traditional energy reserves, the global demand for clean and renewable energy is becoming increasingly urgent. Photovoltaic power generation, as a highly promising clean energy source, occupies an important position in the energy strategies of countries worldwide and is also a key pathway for my country to achieve its "dual-carbon" goals. However, influenced by natural factors such as solar radiation intensity, temperature, and cloud cover changes, photovoltaic power output exhibits significant fluctuations and uncertainties, posing a severe challenge to grid load dispatching, stability maintenance, and peak-shaving and frequency regulation capabilities. Therefore, probabilistic prediction of photovoltaic power ranges has become a crucial basis for ensuring the efficient operation of the power grid, and accurate photovoltaic power probability density prediction can provide more comprehensive risk assessment information for power system planning and dispatching.

[0003] Currently, research methods for probabilistic prediction of photovoltaic power mainly include parametric distribution fitting, nonparametric kernel density estimation, and quantile regression. Parametric methods typically assume that photovoltaic power follows a specific distribution, such as a Beta distribution or a Gaussian mixture distribution. However, due to the complexity of actual power data distributions, this assumption can easily lead to prediction errors. Nonparametric methods, while flexible enough to adapt to different data types, suffer from sensitivity to small samples and high computational complexity. Quantile regression (QR) directly estimates conditional quantiles by minimizing the weighted absolute error, without relying on distribution assumptions. It has good adaptability to asymmetric and multimodal photovoltaic power data; however, traditional QR models struggle to capture complex nonlinear relationships.

[0004] In recent years, with the development of deep learning technology, Quantile Regression Neural Networks (QRNNs) have emerged. They combine the powerful feature extraction capabilities of neural networks with the distributional characterization advantages of quantile regression, overcoming the limitations of traditional QR models while avoiding the shortcomings of parametric and non-parametric methods. However, QRNNs still face two challenges in practical applications: first, quantile crossover and non-differentiability of the loss function may occur during prediction, leading to prediction results that do not meet the monotonicity requirements of the probability distribution and reducing prediction reliability; second, because photovoltaic power output is affected by meteorological factors and exhibits complex fluctuations and uncertainties, a single-structure neural network struggles to simultaneously capture its long-term trends and short-term fluctuations. Summary of the Invention

[0005] This invention aims to address the problems in the aforementioned background technology by providing an adaptive correction method. This method introduces the prediction process of a quantile regression neural network (QRNN), which effectively avoids quantile crossover and ensures the differentiability of the loss function. Simultaneously, it can output prediction results at multiple quantile levels simultaneously, providing support for uncertain prediction. Furthermore, this invention designs a two-stream CNN-BiGRU hybrid neural network architecture, significantly improving the ability to capture temporal features. By combining the quantile regression method with this two-stream hybrid neural network structure, a prediction model with probabilistic output capabilities is constructed.

[0006] To achieve the above objectives, this invention provides a photovoltaic power prediction method based on an adaptive corrected quantile regression neural network, comprising the following steps:

[0007] S1: Obtain historical data on photovoltaic active power and its related influencing factors;

[0008] S2: Preprocess the original dataset;

[0009] S3: Divide the preprocessed dataset and set the neural network parameters;

[0010] S4: The long-term trend features of photovoltaic power data are extracted by combining convolutional neural networks (CNN) with multi-head attention mechanism. At the same time, bidirectional gated recurrent units (BiGRU) are used to learn the short-term fluctuations of photovoltaic power. The two-branch CNN-BiGRU hybrid neural network is combined with multi-output quantile regression to predict conditional quantiles. To avoid the phenomenon of quantile crossover, an adaptive correction mechanism is introduced. In order to solve the problem of non-differentiability of the loss function, a smoothing function is also introduced to approximate the bouncing loss function.

[0011] S5: Train the Adaptive Corrected Quantile Regression Neural Network (ACQRNN) using the training set, and use the trained ACQRNN model to predict on the test set. Calculate various evaluation metrics to assess the prediction performance.

[0012] Further, in step S1, photovoltaic power data from 2018 to 2023 is obtained from the public dataset of the Desert Knowledge Solar Centre (DKASC) in Australia, which includes active power, received effective energy, weather temperature in degrees Celsius, weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, weather daily rainfall, global tilt radiation, and radiative diffuse tilt.

[0013] Furthermore, the data preprocessing described in step S2 includes:

[0014] S21: Missing value checking and imputation;

[0015] Check if the photovoltaic dataset contains missing values; if so, use linear interpolation to fill them in.

[0016] S22: Feature Correlation Analysis;

[0017] The Pearson correlation coefficient method was used to calculate the correlation between photovoltaic active power and various influencing factors. The formula for calculating the Pearson correlation coefficient is as follows: Where cov(·,·) represents the covariance calculation, and N features with low correlation to photovoltaic active power are removed, and the remaining features are input into the ACQRNN model.

[0018] S23: Data normalization;

[0019] The photovoltaic dataset processed in steps S21 and S22 is normalized using the Max-Min method to transform it to the [0,1] interval. The normalization formula is:

[0020] Further, in step S3, the data preprocessed in step S2 is divided into a training set and a test set in an 8:2 ratio. The parameters of the neural network are set as follows: in CNN, the kernel size is set to 3, the pool size is set to 2, and the pool strides are set to 5. In BiGRU, GRU units1 is set to 32, GRU units2 is set to 64, the dropout rate is set to 0.2, the batch size is set to 64, and the learning rate is set to 0.0001.

[0021] Further, in step S4, a dual-stream CNN-BiGRU hybrid neural network architecture is used as the backbone for feature extraction, dividing historical photovoltaic data into long-term and short-term datasets. Prediction accuracy is improved through multi-timescale feature fusion: the CNN branch based on long-term photovoltaic data captures weekly photovoltaic power patterns through two convolutional layers, and performs dimensionality reduction via pooling layers. A multi-head attention mechanism is applied to the CNN branch to highlight features with higher weights in the photovoltaic power data, thus emphasizing more valuable signal features for identification. The BiGRU branch based on short-term photovoltaic power uses bidirectional gated recurrent units to learn daily photovoltaic power time-series dynamics. The features from both branches are concatenated and input into a fully connected layer for high-order nonlinear mapping, ultimately outputting conditional prediction values ​​at multiple quantile levels. The predicted conditional quantiles are obtained using ACQRNN, including the following steps:

[0022] S41: Obtain long-period input data X using a sliding window. Long =[x T-L+1 ,…,x T Short-cycle input data XShort =[x T-S+1 ,…,x T ], predict target Y = y T+1 Input the training set data into the model;

[0023] S42: Constructing a dual-stream CNN-BiGRU hybrid neural network architecture:

[0024] CNNs use one-dimensional convolutional layers to extract information. The one-dimensional convolution operation is described as follows:

[0025]

[0026] Where y t This represents the t-th output value after the convolution operation, where K represents the convolution kernel and length, and w... k x represents the weight of the k-th convolutional kernel. k-t-1 It is represented as the (k-t+1)th value in the input sequence, where ReLU is the activation function.

[0027] The output matrix of the attention mechanism is as follows:

[0028]

[0029] Among them, query key value d k and d v These are the dimensions of the key and the value, respectively; the query and the key have the same dimensions. Each attention head independently calculates its weight and outputs the result, which is then concatenated and linearly transformed to obtain the multi-head attention output.

[0030] MuliHead Attention(Q,K,V)=Concat(head 1 ,…,head h W Q

[0031]

[0032] Among them, the definition and It is a parameter matrix.

[0033] Computing the forget gate r of GRU t Update Gate Z t The hidden layer state h at the previous time step and the current time step t The memory updated at the current moment

[0034] r t =σ(x t Wxz +h t-1 W hz +b z )

[0035] z t =σ(x t W xr +h t-1 W hr +b r )

[0036]

[0037]

[0038] σ and tanh are the activation functions sigmoid and tanh, respectively, and W xz W hz W xr W hr W xh W hh This is the weight matrix. (b) z b r b h This is a bias term.

[0039] BiGRU combines the forward and reverse hidden states and then outputs them. The hidden states of each layer at time t are defined as follows, and the final charge output is h. t The calculation formula is as follows:

[0040]

[0041]

[0042]

[0043] Among them, w t It is the forward propagation weight, v t It is the weight of backpropagation, d t It is a bias term.

[0044] The features extracted from the two branches are fused through a fully connected layer, and the final output layer is:

[0045]

[0046] In the formula W OP and b OP Let represent the weights and corresponding biases of the output layer, respectively, and let Ι(q) be the quantile indicator function (q∈{0.025,…,0.975}).

[0047] S43: Combining a two-stream hybrid neural network with quantile regression, the loss function of the model is obtained as follows:

[0048]

[0049]

[0050] Where N is the total number of samples, and Q is the number of quantile levels to be generated. Therefore, in τ q Load forecast at quantile level It can be derived from the following formula:

[0051]

[0052] S44: To avoid quantile crossover, a quantile difference matrix is ​​defined. The loss function is modified as follows:

[0053]

[0054] Where λ i The quantile adaptive sensing weight is defined as:

[0055]

[0056] in For sample x i The predicted quantile variance reflects local uncertainty. α and β are learnable parameters, with initial values ​​set as α = 0.1 and β = 1.0.

[0057] S45: To address the non-differentiability of the quantile bouncing loss function at the origin, a smoothing function is used to approximate the bouncing loss function, enabling the neural network to better fit the corresponding quantile regression model. The smoothing approximation formula is as follows:

[0058]

[0059] The loss function of the final adaptive corrected quantile regression neural network is:

[0060]

[0061] Further, in step S5, the trained ACQRNN model is applied to the test set to obtain prediction results. The key feature is that the prediction results are evaluated using evaluation metrics. Specifically, the prediction results are evaluated using three metrics: point prediction, interval prediction, and probability density prediction.

[0062] S51: The point prediction evaluation indicators are root mean square error (RMSE), mean absolute error (MAE), and R-squared.

[0063]

[0064]

[0065]

[0066] S52: The evaluation metrics for interval prediction are Interval Coverage Probability (PICP), Average Width of Predicted Interval (PIAW), and Winkler Score (WS).

[0067]

[0068]

[0069] Among them, y i It is the actual value, L i and U i These are the upper and lower bounds of the i-th prediction interval, δ i It is a binary value. It takes the value 1 when the actual value is in the prediction interval, and 0 otherwise. The larger the PICP value, the more actual values ​​are included in the prediction interval and the higher the prediction accuracy.

[0070]

[0071]

[0072]

[0073] S53: The probability density prediction evaluation index is the Continuously Ranked Probability Score (CRPS):

[0074]

[0075]

[0076]

[0077] Where p(y) i ) represents y i The probability density function, It is a probabilistic model derived from predicted values. It is a unit step function.

[0078] Compared with the prior art, the significant advantages of the present invention are reflected in the following aspects:

[0079] 1. To address the shortcomings of traditional models in extracting complex photovoltaic power features, this invention proposes a CNN-BiGRU dual-stream hybrid architecture for scale-based feature extraction, integrating CNN and multi-head attention mechanisms. This architecture effectively enhances the ability to capture temporal features and can output prediction results at multiple quantile levels simultaneously, providing strong support for uncertain prediction.

[0080] 2. Given that current quantile regression models have a simple structure and are difficult to balance accuracy and robustness, this invention combines the quantile regression method with a two-stream hybrid neural network structure to construct a prediction model with probabilistic output capabilities, opening up a new path with great application prospects for complex photovoltaic power prediction.

[0081] 3. To address the issues of quantile crossover and non-differentiability of the loss function at its zero point in quantile regression neural network models, this invention proposes an adaptive loss function correction method and introduces a smoothing function to adjust the final loss function. This method not only ensures the monotonically increasing nature of the predicted quantiles but also guarantees the differentiability of the loss function across the entire domain. Furthermore, this method has broad applicability and can be integrated with numerous neural networks based on shallow and deep learning, without being limited by network structure.

[0082] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0083] Figure 1 This is an overall flowchart of the present invention;

[0084] Figure 2 This is a diagram of the CNN architecture;

[0085] Figure 3 Diagram of the multi-head attention mechanism;

[0086] Figure 4 This is a diagram of the BiGRU structure. Detailed Implementation

[0087] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It should be noted that the embodiments described herein are merely some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on these embodiments without creative effort are within the protection scope of the present invention. Furthermore, the present invention can be applied through other different forms of implementation. Without departing from the basic concept of the present invention, the details in the specification can be adjusted and varied based on different viewpoints and application scenarios. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of this disclosure.

[0088] Example:

[0089] like Figure 1 As shown, this invention provides a photovoltaic power prediction method based on an adaptive corrected quantile regression neural network, comprising the following steps:

[0090] S1: Obtain historical data on photovoltaic active power and its related influencing factors;

[0091] Photovoltaic power data from 2018 to 2023 was obtained from the public dataset of the Desert Knowledge Solar Centre (DKASC) in Australia. This data includes active power, received effective energy, weather temperature in degrees Celsius, weather relative humidity, global horizontal radiation, diffuse horizontal radiation, wind direction, daily rainfall, global tilt radiation, and radiative diffuse tilt.

[0092] S2: Preprocess the original dataset;

[0093] S21: Missing value checking and imputation;

[0094] Check if the photovoltaic dataset contains missing values; if so, use linear interpolation to fill them in.

[0095] S22: Feature Correlation Analysis;

[0096] The Pearson correlation coefficient method was used to calculate the correlation between photovoltaic active power and various influencing factors. The formula for calculating the Pearson correlation coefficient is as follows: Where cov(·,·) represents the covariance calculation, and N features with low correlation to photovoltaic active power are removed, and the remaining features are input into the ACQRNN model.

[0097] S23: Data normalization;

[0098] The photovoltaic dataset processed in steps S21 and S22 is normalized using the Max-Min method to transform it to the [0,1] interval. The normalization formula is:

[0099] S3: Divide the preprocessed dataset and set the neural network parameters;

[0100] The data preprocessed in step S2 is divided into a training set and a test set in an 8:2 ratio. The parameters of the neural network are set as follows: in CNN, the kernel size is set to 3, the pool size is set to 2, and the pool strides are set to 5. In BiGRU, GRU units1 is set to 32, GRU units2 is set to 64, the dropout rate is set to 0.2, the batch size is set to 64, and the learning rate is set to 0.0001.

[0101] S4: Construct the ACQRNN model:

[0102] A dual-stream CNN-BiGRU hybrid neural network architecture was used as the backbone for feature extraction, dividing historical photovoltaic (PV) data into long-term and short-term datasets. Prediction accuracy was improved through multi-timescale feature fusion: the CNN branch based on long-term PV data captured weekly PV power patterns through two convolutional layers, followed by dimensionality reduction via pooling layers. A multi-head attention mechanism was applied to the CNN branch to highlight features with higher weights in the PV power data, thus emphasizing more valuable signal features for identification. The BiGRU branch based on short-term PV power used bidirectional gated recurrent units to learn daily PV power time-series dynamics. The features from both branches were concatenated and input into a fully connected layer for high-order nonlinear mapping, ultimately outputting conditional prediction values ​​at multiple quantile levels. The predicted conditional quantiles were obtained using ACQRNN, including the following steps:

[0103] S41: Obtain long-period input data X using a sliding window. Long =[x T-L+1 ,…,x T Short-cycle input data X Short =[x T-S+1 ,…,x T ], predict target Y = y T+1 Input the training set data into the model;

[0104] S42: Constructing a dual-stream CNN-BiGRU hybrid neural network architecture:

[0105] CNNs use one-dimensional convolutional layers to extract information. The one-dimensional convolution operation is described as follows:

[0106]

[0107] Where y t This represents the t-th output value after the convolution operation, where K represents the convolution kernel and length, and w... k x represents the weight of the k-th convolutional kernel. k-t-1 It is represented as the (k-t+1)th value in the input sequence, where ReLU is the activation function.

[0108] The output matrix of the attention mechanism is as follows:

[0109]

[0110] Among them, query key value d k and d v These are the dimensions of the key and the value, respectively; the query and the key have the same dimensions. Each attention head independently calculates its weight and outputs the result, which is then concatenated and linearly transformed to obtain the multi-head attention output.

[0111] MuliHead Attention(Q,K,V)=Concat(head 1 ,…,head h W Q

[0112]

[0113] Among them, the definition and It is a parameter matrix.

[0114] Computing the forget gate r of GRU t Update Gate Z t The hidden layer state h at the previous time step and the current time step t The memory updated at the current moment

[0115] r t =σ(x t W xz +h t-1 W hz +b z )

[0116] z t =σ(x t W xr +h t-1 W hr +b r )

[0117]

[0118]

[0119] σ and tanh are the activation functions sigmoid and tanh, respectively, and W xz W hz W xr W hr W xh W hh This is the weight matrix. (b) z b r b h This is a bias term.

[0120] BiGRU combines the forward and reverse hidden states and then outputs them. The hidden states of each layer at time t are defined as follows, and the final charge output is h. t The calculation formula is as follows:

[0121]

[0122]

[0123]

[0124] Among them, w t It is the forward propagation weight, v t It is the weight of backpropagation, d t It is a bias term.

[0125] The features extracted from the two branches are fused through a fully connected layer, and the final output layer is:

[0126]

[0127] In the formula W OP and b OP Let represent the weights and corresponding biases of the output layer, respectively, and let Ι(q) be the quantile indicator function (q∈{0.025,…,0.975}).

[0128] S43: Combining a two-stream hybrid neural network with quantile regression, the loss function of the model is obtained as follows:

[0129]

[0130]

[0131] Where N is the total number of samples, and Q is the number of quantile levels to be generated. Therefore, in τ q Load forecast at quantile level It can be derived from the following formula:

[0132]

[0133] S44: To avoid quantile crossover, a quantile difference matrix is ​​defined. The loss function is modified as follows:

[0134]

[0135] Where λ i The quantile adaptive sensing weight is defined as:

[0136]

[0137] in For sample x i The predicted quantile variance reflects local uncertainty. α and β are learnable parameters, with initial values ​​set as α = 0.1 and β = 1.0.

[0138] S45: To address the non-differentiability of the quantile bouncing loss function at the origin, a smoothing function is used to approximate the bouncing loss function, enabling the neural network to better fit the corresponding quantile regression model. The smoothing approximation formula is as follows:

[0139]

[0140] The loss function of the final adaptive corrected quantile regression neural network is:

[0141]

[0142] S5: The trained ACQRNN model is applied to the test set to obtain prediction results. The key feature is that the prediction results are evaluated using evaluation metrics. Specifically, the prediction results are evaluated using three metrics: point prediction, interval prediction, and probability density prediction.

[0143] S51: The point prediction evaluation indicators are root mean square error (RMSE), mean absolute error (MAE), and R-squared.

[0144]

[0145]

[0146]

[0147] S52: The evaluation metrics for interval prediction are Interval Coverage Probability (PICP), Average Width of Predicted Interval (PIAW), and Winkler Score (WS).

[0148]

[0149]

[0150] Among them, y i It is the actual value, L i and U i These are the upper and lower bounds of the i-th prediction interval, δ i It is a binary value. It takes the value 1 when the actual value is in the prediction interval, and 0 otherwise. The larger the PICP value, the more actual values ​​are included in the prediction interval and the higher the prediction accuracy.

[0151]

[0152]

[0153]

[0154] S53: The probability density prediction evaluation index is the Continuously Ranked Probability Score (CRPS):

[0155]

[0156]

[0157]

[0158] Where p(y) i ) represents y i The probability density function, It is a probabilistic model derived from predicted values. It is a unit step function.

[0159] This invention addresses the performance limitations of traditional models in extracting complex photovoltaic power features by proposing a CNN-BiGRU dual-stream hybrid architecture for scaled feature extraction. By fusing CNN with a multi-head attention mechanism, it effectively enhances the ability to capture temporal features and supports the simultaneous output of multi-level quantile prediction results, enabling predictions for uncertainties. Furthermore, addressing the issue of current quantile regression models having a single structure and difficulty in balancing accuracy and robustness, this invention combines quantile regression with a dual-stream hybrid neural network structure to construct a prediction model with probabilistic output capabilities, providing a new path for complex photovoltaic power prediction. Finally, to address the quantile crossover and loss function non-differentiability issues in quantile regression neural networks, an adaptive loss function correction method is proposed. By introducing a smoothing function to adjust the loss function, it ensures both the monotonically increasing nature of the predicted quantiles and the global differentiability of the loss function. This method is also widely applicable, can be integrated with various shallow and deep learning networks, and is not limited by network structure.

[0160] The embodiments of the present invention have been described in detail above. For those skilled in the art, after understanding the basic principles of the present invention, various modifications, equivalent substitutions, improvements, etc., can be made to the present invention without departing from the concept of the present invention, and these should all be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on an adaptive corrected quantile regression neural network, characterized in that, Includes the following steps: S1: Obtain historical data on photovoltaic active power and its related influencing factors; S2: Preprocess the original dataset; S3: Divide the preprocessed dataset and set the neural network parameters; S4: The long-term trend features of photovoltaic power data are extracted by combining convolutional neural networks (CNN) with multi-head attention mechanism. At the same time, bidirectional gated recurrent units (BiGRU) are used to learn the short-term fluctuations of photovoltaic power. The two-branch CNN-BiGRU hybrid neural network is combined with multi-output quantile regression to predict conditional quantiles. To avoid the phenomenon of quantile crossover, an adaptive correction mechanism is introduced. In order to solve the problem of non-differentiability of the loss function, a smoothing function is also introduced to approximate the bouncing loss function. S5: Train the Adaptive Corrected Quantile Regression Neural Network (ACQRNN) using the training set, and use the trained ACQRNN model to predict on the test set. Calculate various evaluation metrics to assess the prediction performance.

2. The photovoltaic power prediction method using an adaptive modified quantile regression neural network according to claim 1, characterized in that, In step S1, photovoltaic power data and related influencing factor data are obtained from the public dataset of the Desert Knowledge Solar Centre (DKASC) in Australia.

3. The photovoltaic power prediction method using an adaptive corrected quantile regression neural network according to claim 1, characterized in that, The data preprocessing in step S2 includes: S21: Missing value checking and imputation; Check if the photovoltaic dataset contains missing values; if so, use linear interpolation to fill them in. S22: Feature Correlation Analysis; The Pearson correlation coefficient method was used to calculate the correlation between photovoltaic active power and various influencing factors. The formula for calculating the Pearson correlation coefficient is as follows: Where cov(·,·) represents the covariance calculation, and N features with low correlation to photovoltaic active power are removed, and the remaining features are input into the ACQRNN model; S23: Data normalization; The photovoltaic dataset processed in steps S21 and S22 is normalized using the Max-Min method to transform it into the [0,1] interval. The normalization formula is:

4. The photovoltaic power prediction method using an adaptive modified quantile regression neural network according to claim 1, characterized in that, In step S3, the data preprocessed in step S2 is divided into a training set and a test set in an 8:2 ratio.

5. The photovoltaic power prediction method using an adaptive corrected quantile regression neural network according to claim 1, characterized in that, Step S4 describes using a dual-stream CNN-BiGRU architecture as the feature extraction backbone, dividing historical photovoltaic data into long-term and short-term datasets, and fusing features across multiple time scales to improve prediction accuracy: the CNN branch based on long-term data uses two convolutional modules to capture weekly power patterns, and after dimensionality reduction by pooling layers, a multi-head attention mechanism is applied to enhance the weights of key features; the BiGRU branch based on short-term data uses bidirectional gated recurrent units to learn daily time-series dynamics, and the features from the two branches are concatenated and then subjected to high-order nonlinear mapping through a fully connected layer to output multi-level conditional prediction quantiles. The ACQRNN then obtains the predicted conditional quantiles. The detailed process of step S4 is as follows: S41: Obtain long-period input data X using a sliding window. Long =[x T-L+1 ,…,x T Short-cycle input data X Short =[x T-S+1 ,…,x T ], predict target Y = y T+1 Input the training set data into the model; S42: CNNs use one-dimensional convolutional layers to extract information. The one-dimensional convolution operation is described as follows: Where y t This represents the t-th output value after the convolution operation, where K represents the convolution kernel and length, and w... k Let x represent the weight of the k-th convolutional kernel. k-t-1 Represented as the (k-t+1)th value in the input sequence, where ReLU is the activation function; The output matrix of the attention mechanism is as follows: Among them, query key value d k and d v These are the dimensions of the key and the value, respectively. The query and the key have the same dimensions. The attention heads independently calculate their weights and output the results. Finally, the multi-head attention output is obtained by concatenation and linear transformation. MuliHead Attention(Q,K,V)=Concat(head 1 ,…,head h )W Q Among them, the definition and For parameter matrices; Computing the forget gate r of GRU t Update Gate Z t The hidden layer state h at the previous time step and the current time step t The memory updated at the current moment r t =σ(x t W xz +h t-1 W hz +b z ) z t =σ(x t W xr +h t-1 W hr +b r ) σ and tanh are the activation functions sigmoid and tanh, respectively, and W xz W hz W xr W hr W xh W hh Let b be the weight matrix. z b r b h For bias terms; BiGRU combines the forward and reverse hidden states and then outputs them. The hidden states of each layer at time t are as follows, and the final charge output is h. t The calculation formula is as follows: Among them, w t It is the forward propagation weight, v t It is the weight of backpropagation, d t It is a bias term; The features extracted from the two branches are fused through a fully connected layer, and the final output layer is: In the formula W OP and b OP Let represent the weights and corresponding biases of the output layer, respectively, and let Ι(q) be the quantile indicator function (q∈{0.025,…,0.975}). S43: Combining a two-stream hybrid neural network with quantile regression, the loss function of the model is obtained as follows: Where N is the total number of samples, Q is the number of quantile levels to be generated, and so on, in τ q Load forecast at quantile level It can be derived from the following formula: S44: To avoid quantile crossover, a quantile difference matrix is ​​defined. The loss function is modified as follows: Where λ i The quantile adaptive sensing weight is defined as: in For sample x i The predicted quantile variance reflects local uncertainty. α, β are learnable parameters, with initial values ​​set as α = 0.1 and β = 1.

0. S45: To address the non-differentiability of the quantile bouncing loss function at the origin, a smoothing function is used to approximate the bouncing loss function, enabling the neural network to better fit the corresponding quantile regression model. The smoothing approximation formula is as follows: The loss function of the final adaptive corrected quantile regression neural network is:

6. The photovoltaic power prediction method using an adaptive corrected quantile regression neural network according to claim 1, characterized in that, Step S5, which involves applying the trained ACQRNN model to the test set to obtain prediction results, is characterized by evaluating the prediction results using evaluation metrics. Specifically, the evaluation is conducted using three metrics: point prediction, interval prediction, and probability density prediction. S51: The point prediction evaluation indicators are root mean square error (RMSE), mean absolute error (MAE), and R-squared. S52: The evaluation metrics for interval prediction are Interval Coverage Probability (PICP), Average Width of Predicted Interval (PIAW), and Winkler Score (WS). Among them, y i It is the actual value, L i and U i These are the upper and lower bounds of the i-th prediction interval, respectively, δ i It is a binary value. It takes 1 when the actual value is in the prediction interval, and 0 otherwise. The larger the PICP value, the more actual values ​​are included in the prediction interval and the higher the prediction accuracy. S53: The probability density prediction evaluation index is the Continuously Ranked Probability Score (CRPS): Where p(y i ) represents y i The probability density function, It is a probabilistic model derived from predicted values. It is a unit step function.

Images