Method for predicting net load in shaguo barren region under sandstorm weather based on rpinn

Abstract

The present application relates to the field of meteorological electric power prediction, and discloses a dust storm weather-based Sha Gao Barren region net load prediction method based on RPINN, first, aiming at the accurate net load prediction demand of the power grid optimal dispatching of the Sha Gao Barren region under the extreme weather of dust storm, a targeted net load prediction framework is constructed to improve its support ability for dispatching decision; secondly, the generalized partial differential equation is introduced to construct the load prediction model embedded with physical constraints, and a robust physical information neural network is constructed to enhance the learning robustness of the model under small sample conditions; the Mamba-SE algorithm fusing the SE mechanism is proposed to improve the efficiency of the physical constraint and extract key time sequence feature information from the high volatility and high uncertainty input under complex disturbance. The present application can realize high-precision prediction of net load under small sample conditions and complex meteorological disturbance, provide reliable decision support for the power grid optimal dispatching of the Sha Gao Barren region, and thus achieve good prediction performance.

Description

Technical Field

[0001] This invention relates to the field of meteorological electrical power prediction, specifically a method for predicting net load in desert areas under sandstorm weather based on RPINN. Background Technology

[0002] With the accelerated construction of new power systems, the proportion of new energy sources such as wind power and photovoltaics in the power supply structure continues to rise. my country's desert and Gobi regions, relying on their abundant wind and solar resources, have gradually developed into strategic areas for new energy development. However, the frequent sandstorms in these areas have significant and complex disturbance effects on new energy output and surrounding loads: photovoltaic power generation experiences a sharp drop in output due to sand and dust covering the components; wind turbines are forced to shut down due to sandstorm protection and safety requirements; and rapid changes in weather conditions can also cause abnormal fluctuations in power load. These factors collectively lead to a significant increase in the risks of power grid dispatching and operation. Therefore, achieving accurate and reliable net load forecasting under sandstorm conditions is crucial for improving the stability of new energy transmission from desert and Gobi regions.

[0003] Currently, this forecasting task still faces two prominent challenges: First, historical samples of extreme weather events are extremely scarce, making it difficult to adequately support the effective training of forecasting models; second, dust storms are characterized by their long duration, wide impact range, and complex and variable evolution, resulting in intricate disturbances to both the power supply and load sides, leading to strong volatility and high uncertainty in net load data. Therefore, research on net load forecasting methods for sandstorm-affected desert areas is urgently needed to address these two key issues.

[0004] To address the issues of data scarcity and high volatility, incorporating Physics-Informed Neural Networks (PINNs) into dust storm net load prediction is a feasible technical approach. By embedding the partial differential equations describing the power load response as physical constraints into the loss function, the model's dependence on large-scale training data can be reduced. However, under extreme dust storm conditions, theoretical physical formulas (such as photovoltaic irradiance models and wind power curves) deviate significantly from actual output due to non-ideal factors like dust shading and blade erosion. This leads to frequent directional conflicts between the physical consistency loss gradient and the data fitting loss gradient during training. Such gradient conflicts cause performance degradation after the model reaches a stage optimum—physical constraints "pull" the model away from the local optimum of data fitting, causing training to oscillate or even diverge. Existing PINN improvement schemes often employ gradient harmonization or reweighting strategies, attempting to find a compromise balance among conflicting gradients. However, in extreme scenarios with significant physical deviations, harmonization strategies cannot fundamentally eliminate conflicts, and the model still faces insufficient learning robustness. Therefore, how to construct a robust learning mechanism that can adaptively manage physical constraints and data fitting under extreme conditions has become a key challenge in predicting net load physical information during sandstorms. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to provide a net load forecasting method for desert and Gobi areas under sandstorm weather based on RPINN (Robust Physics-Informed Neural Network). This method can achieve high-precision net load forecasting under small sample conditions and complex meteorological disturbances, providing reliable decision support for the optimized scheduling of the power grid in desert and Gobi areas. The technical solution is as follows:

[0006] A method for predicting net load in the desert region under dust storm weather based on RPINN includes the following steps:

[0007] Step 1: To address the lack of historical samples under extreme weather conditions such as sandstorms, Pearson correlation coefficient is used to analyze the correlation between photovoltaic, wind power, and load data and meteorological data under extreme weather conditions such as sandstorms. The obtained highly correlated data sets are input into the time-based generative adversarial network model for data augmentation. The generated data with the maximum mean difference less than a set threshold are selected and combined with the original data to form an enhanced dataset, which is used as the input for the subsequent prediction network.

[0008] Step 2: Based on the correlation analysis results of the load data and meteorological data obtained in Step 1, a generalized partial differential equation for load response relationship modeling is constructed. The coefficients of the generalized partial differential equation are adaptively solved using the Physical Information Neural Network (PINN) to obtain a power load response relationship model with definite coefficients, so as to construct the physical constraints for load forecasting under sandstorm weather. Among them, the difference between the partial derivatives of power with respect to key environmental factors is used as the physical consistency loss.

[0009] Step 3: After expanding the feature dimension of the input data through the encoding layer, the feature sequence is encoded and then enters the Mamba-SE module. After pooling and decoding operations, the final prediction result is output. The Mamba-SE module introduces a compression and activation channel attention mechanism before the pooling and decoding of the Mamba algorithm to selectively recalibrate the temporal feature channels.

[0010] Step 4: During model training, the normalized synthetic gradient norm is used as the conflict metric. When the conflict metric after sliding window averaging exceeds the preset threshold, the physical consistency loss is stopped from participating in the model parameter update, and the model training is driven only by the data fitting loss; otherwise, the network parameters are updated by the weighted sum of the data fitting loss and the physical consistency loss. Through repeated iterative training, the net load of the desert area under the extreme weather conditions of sandstorms is finally accurately predicted.

[0011] The beneficial effects of this invention are:

[0012] 1) This invention constructs a load forecasting model with embedded physical constraints by introducing generalized partial differential equations, and then embeds the physical consistency loss into the net load forecasting model training process to reduce the model's dependence on large-scale training data. Experimental results show that the method of this invention exhibits good adaptability to the characteristics of small-scale data samples under extreme weather conditions such as sandstorms. Compared with the comparison methods, the method of this invention maintains the lowest prediction error under various data scales: taking the 24-hour net load forecasting task as an example, when data is sufficient, the accuracy of the method of this invention is improved by 34.39%, 16.51%, and 23.35% respectively compared with the single Mamba-SE, the RPINN only, and the TimeGan only scheme, indicating that the method of this invention can make full use of high-quality generated samples to obtain higher prediction accuracy under high data volume; when data is scarce, the method of this invention achieves significant error reductions of 51.46%, 20.31%, and 23.51% compared with the single Mamba-SE, the collaborative optimization scheme, and the TimeGan only scheme, proving that it can still maintain the best prediction accuracy when the sample is extremely scarce.

[0013] 2) This invention constructs a robust physical information neural network based on conflict metrics, enabling the model to exhibit more robust learning performance than the traditional PINN under extreme conditions. Experimental results show that this method effectively overcomes the potential performance degradation of the traditional PINN after the optimal round under extreme conditions, maintaining a downward trend after the error reaches a stage minimum and continuously converging to a lower error level. In experiments, this method achieves higher learning robustness and prediction accuracy in load and wind power forecasting, thus significantly outperforming the traditional PINN and purely data-driven models in overall net load forecasting.

[0014] 3) This invention proposes an RPINN-Mamba-SE model that integrates a robust physical information neural network and an SE channel attention mechanism to address the problems of weak model learning robustness and difficulty in extracting key temporal features caused by small sample data and complex perturbations under extreme weather conditions such as sandstorms. Ultimately, it achieves high-precision prediction of net load in desert and Gobi areas under sandstorm weather. In the net load prediction experiment, the model of this invention achieved the best performance at both scales, with net load prediction errors of 1.694% and 1.888%, respectively. The 24-hour net load prediction error was reduced by 28.40%, 29.61%, and 43.28% compared to the RPINN-Mamba model, the Mamba-SE model, and the traditional Mamba model, respectively. Moreover, its improvement over the traditional Mamba model is significantly better than the gain brought by the introduction of a single module, verifying the complementary advantages of the robust physical mechanism and the SE module. Attached Figure Description

[0015] Figure 1 This is a diagram of the physical information neural network algorithm framework of the present invention.

[0016] Figure 2 This is a framework diagram of the net load prediction algorithm based on a robust physical information neural network of the present invention.

[0017] Figure 3 This is a diagram illustrating the framework of the Mamba-SE prediction algorithm based on the fusion channel attention mechanism of this invention.

[0018] Figure 4 This is a flowchart of the net load prediction process for the desert-golan region under extreme sandstorm weather based on a robust physical information neural network, as described in this invention; in the figure, MSE represents the mean square error. The gradient of the physical consistency loss. The gradient of the data fitting loss is denoted by ; SILU is the activation function of the Sigmoid Linear Unit. It should be noted that, in order to visually distinguish data structures and processing modules in the diagram, the legends divided into small blocks (such as meteorological and load data sequences, time-series feature tensors, etc.) represent multi-dimensional data vectors, matrices, or feature tensors; while single solid-line boxes and rounded rectangles without internal division (such as encoded input, physical consistency loss, etc.) represent single data scalars, neural network layers, logical operations, or functional modules.

[0019] Figure 5 This is a heatmap of Pearson correlation analysis for this invention.

[0020] Figure 6 This is the residual histogram of the power load response relationship model of this invention.

[0021] Figure 7 This is a comparison chart of the 24-hour net load prediction results of each model under different proportions of the original dataset in this invention.

[0022] Figure 8 This is a comparison chart of the 72-hour net load prediction results of each model under different proportions of the original dataset in this invention.

[0023] Figure 9 This is a comparison chart of the model training process under different physical constraint strategies of the present invention—a 24-hour load prediction chart.

[0024] Figure 10 This is a comparison chart of the model training process under different physical constraint strategies of the present invention—a 72-hour load prediction chart.

[0025] Figure 11 This is a comparison diagram of the model training process under different physical constraint strategies of the present invention—a wind power prediction diagram.

[0026] Figure 12 This is a comparison chart of the photovoltaic prediction results of the 24-hour net load system under sandstorm weather conditions according to the present invention.

[0027] Figure 13 This is a comparison chart of the wind power prediction results of the 24-hour net load system under sandstorm weather conditions according to the present invention.

[0028] Figure 14 This is a comparison chart of the 24-hour net load prediction results of the system under sandstorm weather conditions according to the present invention.

[0029] Figure 15 This is a comparison chart of the net load prediction results of the 24-hour net load system under sandstorm weather conditions according to the present invention.

[0030] Figure 16This is a comparison chart of the photovoltaic prediction results of the 72-hour net load system under sandstorm weather conditions according to the present invention.

[0031] Figure 17 This is a comparison chart of the wind power prediction results of the 72-hour net load system under sandstorm weather conditions according to the present invention.

[0032] Figure 18 This is a comparison chart of the 72-hour net load prediction results of the system under sandstorm weather conditions according to the present invention.

[0033] Figure 19 This is a comparison chart of the net load prediction results of the 72-hour net load system under sandstorm weather conditions according to the present invention. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0035] Reference Figures 1 to 4 As an embodiment of the present invention, a method for predicting net load in desert areas under sandstorm weather based on RPINN is provided, including:

[0036] S1: To address the lack of historical samples during extreme sandstorm weather, Pearson correlation coefficient analysis was used to analyze the correlation between photovoltaic, wind power, and load data and meteorological data during extreme sandstorm weather. Subsequently, the obtained highly correlated data sets were input into the TimeGan (Time Generative Adversarial Network) model for data augmentation. Generated data with an MMD (Maximum Mean Discrepancy) less than a set threshold (THRESH) were selected and combined with the original data to form an enhanced dataset, which was then used as input for the subsequent prediction network.

[0037] Specifically, considering the limited sample size of sandstorm processes, this invention argues that using correlation analysis methods more sensitive to sample size, such as MIC, for feature selection might introduce estimation bias and reduce the stability of conclusions. Therefore, this invention uses the Pearson correlation coefficient to measure the strength of linear correlation, obtaining relatively robust correlation assessment results under small sample conditions. The formula for calculating the Pearson correlation coefficient is as follows:

[0038] ;

[0039] In the formula: , For the first Each observed object in the variable , The value of ; , The mean of the sample; Indicates the sample length of the time series; For variables , The Pearson correlation coefficient was calculated. To avoid information leakage from the test set, the correlation analysis was performed only based on the training set samples; the test set was not involved in the correlation coefficient calculation or feature selection process.

[0040] After feature selection, to further address the severe shortage of training samples under extreme weather conditions like sandstorms, this invention introduces TimeGan for data augmentation before predictive modeling. TimeGan combines adversarial learning from GANs with temporal modeling, introducing latent space-based supervision constraints. This ensures that the generated sequences not only maintain consistency with real data in terms of statistical distribution but also better preserve temporal dependency structures. Since sandstorms cause complex disturbances to power grid operation, resulting in highly volatile and uncertain net load data, TimeGan can more effectively learn typical evolutionary patterns compared to GANs. This generates synthetic sequences consistent with the statistical characteristics of real sandstorm scenarios, enhancing the model's learning and generalization capabilities for extreme scenarios.

[0041] To verify that the generated data and the real data have the same statistical distribution and to ensure the usability of the generated data, this invention uses the Maximum Mean Discrepancy (MMD) as a metric. The generated data is only included in subsequent tasks if the MMD between the generated data and the real data is less than a pre-set threshold.

[0042] Let the real sample set be The generated sample set is In kernel function Below, the maximum mean difference is:

[0043] ;

[0044] In the formula: The maximum mean difference between the real sample set X and the generated sample set Y; and The sample size for the real sample set and the generated sample set; Take the Gaussian radial basis function kernel.

[0045] S2: Based on the correlation analysis results of the load data and meteorological data obtained in S1, a generalized partial differential equation for load response relationship modeling is constructed, and the coefficients of the equation are solved with the help of PINN to obtain a power load response relationship model with definite coefficients, so as to construct the physical constraints for load forecasting under sandstorm weather.

[0046] Specifically, when constructing physical constraints for training photovoltaic prediction models, one common approach is to directly use the difference between predicted and theoretical power as the physical constraint. However, in extreme weather scenarios, the deviation between theoretical and actual power output is usually large, and this method reduces the stability of the model training process. Therefore, this invention introduces the difference in partial derivatives of electrical power with respect to key environmental factors as a physical consistency loss. This trend-based relationship is more robust to uncertainties in the external environment and deviations in system parameter settings, exhibits relatively smaller deviations from the actual physical process, and effectively avoids information redundancy, thus stimulating the algorithm's potential in uncovering deeper physical mechanisms.

[0047] To fully utilize solar energy and improve operating efficiency, photovoltaic (PV) generator sets typically employ maximum power point tracking (MPPT) technology to control solar cell power generation in practical applications. In this case, the PV generator set's power output can be calculated using the following formula:

[0048] ;

[0049] In the formula: , These represent the number of solar cell strings and the number of parallel connections in the photovoltaic unit. and These represent the maximum operating current and maximum operating voltage of the solar cell, respectively.

[0050] The output power of a solar cell depends not only on its electrical parameters but is also significantly affected by environmental conditions, primarily temperature and solar irradiance. Therefore, it is necessary to determine the maximum operating current under standard operating conditions. With maximum operating point voltage After introducing a compensation factor correction, the corrected operating point current and operating point voltage are as follows:

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] In the formula: , These represent irradiance and temperature, respectively. , These are standard irradiation intensity and standard temperature, respectively. , , These are empirical coefficients, typically taken as 0.0025, 0.5, and 0.00288 respectively; The corrected operating point current, This is the corrected operating point voltage; , These are the deviations between the actual temperature and the standard temperature, and the relative deviations between the actual irradiance and the standard irradiance, respectively. It is a natural constant.

[0056] From the above formula, the first derivative of photovoltaic power generation with respect to solar irradiance can be calculated using the following formula:

[0057] ;

[0058] In terms of wind power generation forecasting, sandstorms can severely corrode wind turbines: on the one hand, they exacerbate blade wear, increase surface roughness, weaken blade mechanical properties, and reduce unit operating efficiency; on the other hand, fine sand easily enters rotating parts and oil seals, accelerating wear and compromising sealing. Simultaneously, severe sandstorm environments significantly increase the safety risks of wind farm operation and maintenance, thus often requiring shutdowns several hours after a sandstorm. Therefore, the power generation of a single wind turbine under extreme weather conditions during a sandstorm is... Wind turbine generator shutdown state control function and wind turbine generator power generation They are respectively:

[0059] ;

[0060] ;

[0061] ;

[0062] In the formula: The power generation capacity of the wind turbine generator set, The power generation capacity of a single wind turbine. air density, For wind speed, The radius of the swept surface of the wind turbine; Power coefficient; This refers to the number of wind turbines within the wind turbine generator set. Indicates the time when the sandstorm occurred. This indicates the delay in wind farm shutdown after a disaster, usually expressed as 4 hours. , and These are the cut-in wind speed, cut-out wind speed, and rated wind speed of the wind turbine generator set, respectively. The rated power output of a single wind turbine. This is a control function for the shutdown status of wind turbine generators under extreme weather conditions such as sandstorms. A value of 1 indicates normal operation, and a value of 0 indicates shutdown due to disaster. This is the differential symbol.

[0063] Similar to the physical constraint equations for photovoltaic generator sets, this invention introduces the difference between the first derivatives of wind power generation with respect to wind speed into the model to construct a physical consistency loss. The formula for calculating the first derivative of wind power generation with respect to wind speed is as follows:

[0064] ;

[0065] ;

[0066] In terms of load forecasting, inspired by differential polynomial neural networks, this invention introduces a class of generalized partial differential equations that can characterize the input-output response relationship of a system based on data samples. Their general mathematical form is as follows:

[0067] ;

[0068] in, Let be the coefficients of the terms in the generalized partial differential equation. Let be the unknown function being fitted, representing the system state or output. is the independent variable or the system's input variable; D is the number of independent variables affecting load changes.

[0069] Existing research shows that generalized partial differential equations can effectively characterize the unknown response relationships hidden in the observed samples, and have demonstrated good modeling ability and prediction accuracy in load prediction tasks based on differential multinomial neural networks.

[0070] Based on this, the present invention uses the above equation to construct the load response relationship equation: At this time, the unknown function Represents the power system load, input variables The various factors affecting load changes are represented, including time-coded and meteorological factors screened by correlation. Then, the coefficients of the equation are adaptively solved using PINN to obtain a load response relationship equation with definite coefficients, which is used to characterize the dynamic response of the load to multiple factors.

[0071] It should be noted that electricity load is affected by a variety of factors and is difficult to characterize rigorously by explicit physical equations. Therefore, the generalized partial differential equations introduced in this invention are mainly used to identify the response relationship of load to time and meteorological variables, and to constrain the model prediction results to conform to the identified load response law, thereby improving the physical consistency between the prediction results and the actual load change law.

[0072] S3: After the feature dimension of the input data is expanded by the encoding layer, the feature sequence enters the Mamba-SE module after encoding. After pooling and decoding operations, the final prediction result is output. The Mamba-SE module introduces a compression and excitation channel attention mechanism before the pooling and decoding of the Mamba algorithm to selectively recalibrate the temporal feature channels.

[0073] The Mamba-SE module introduces an SE (Squeeze and Excitation) channel attention mechanism before pooling and decoding in the Mamba algorithm to selectively recalibrate temporal feature channels.

[0074] Specifically, considering the significant differences in the variable objects, mechanistic meanings, and mathematical forms of the physical constraint equations for photovoltaic, wind power, and load mentioned above, this invention employs an indirect method to predict the system's net load in order to facilitate the design of a loss function that incorporates physical constraints. Three independent neural network sub-models are constructed for photovoltaic power generation prediction, wind power generation prediction, and load prediction, respectively, and the difference between the load prediction result and the photovoltaic and wind power prediction results is used as the system's net load prediction result. The loss functions of these three sub-models can then be expressed as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula: , and These represent the total losses of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and electricity load forecasting, respectively. , and These represent the data fitting losses for photovoltaic power generation forecasts, wind power generation forecasts, and electricity load forecasts, respectively. , and The physical consistency loss of photovoltaic power generation forecasting, wind power generation forecasting and electricity load forecasting; , , These represent the weighting coefficients of the data fitting loss and physical consistency loss of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and power load forecasting, respectively.

[0079] When constructing the data fitting loss, since the training objective of the data-driven parts of the three sub-networks is consistent—namely, minimizing the difference between the predicted and actual values—the mean squared error between the predicted and actual values ​​is used as a unified data loss function. Therefore, the data fitting losses of the three sub-models are as follows:

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula: , and represents the data fitting loss of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and electricity load forecasting, respectively; M is the number of training sample sequence groups for the model. , , The first generation of the three neural network sub-models The true output power value of each training sample sequence; , , The first generation of the three neural network sub-models The predicted output power of each training sample sequence.

[0084] When constructing the physical consistency loss, this invention embeds various established physical constraint equations into the corresponding sub-networks. Since the forms of the physical constraint equations on the source side and the load side differ to some extent, it is necessary to construct the physical consistency loss separately.

[0085] For the photovoltaic and wind power prediction models, after calculating the theoretical values ​​of the first-order partial derivatives of photovoltaic power generation with respect to irradiance and wind power generation with respect to wind speed based on the above formula, the physical consistency loss of the photovoltaic and wind power prediction models can be constructed, namely:

[0086] ;

[0087] ;

[0088] In the formula: and The measured irradiance and measured wind speed are from the model database; these are the measured data. and The theoretical output is calculated from measured data; , The output power prediction values ​​generated by the neural network sub-models for photovoltaic power generation prediction and wind power generation prediction.

[0089] Figure 1 middle, P true (i) represents the true output power value of the i-th training sample sequence. P pred (i) represents the predicted output power of the i-th training sample sequence.

[0090] For load forecasting models, electricity load forecasting data Measured data of various factors affecting load changes , After substituting the coefficients into the following equation and solving for them using PINN, the resulting deterministic load response equation is obtained by substituting the power load forecast data. Based on the above measured data, the squared value of the equation residuals can be directly used to construct the physical consistency loss of the load forecasting model:

[0091] ;

[0092] In the formula: This represents the residual value of the equation calculated by substituting the measured data and predicted data at the corresponding time into the generalized partial differential equation; Measured data for various factors affecting load changes.

[0093] Building upon the loss function, this invention further introduces a physical constraint exit mechanism based on conflict metrics to effectively mitigate the negative impact of errors in the physical constraint equations on the training process and ensure the stability of model training.

[0094] During model training using the traditional PINN algorithm, errors in the physical constraint equations can cause significant conflicts between the physical consistency loss and the data fitting loss during model updates. Inspired by the GradNorm algorithm, this invention uses the normalized synthetic gradient norm after window averaging as a conflict metric to determine the degree of conflict. When the conflict metric exceeds a certain threshold during model learning, the physical consistency loss stops participating in model updates, thereby avoiding excessive interference from constraint equation errors in the training process.

[0095] Specifically, for each prediction model, at the 1st Rounds, recorded The gradients of the data fitting loss and the physical consistency loss over the training sample sequences are: and :

[0096] ;

[0097] ;

[0098] In the formula: For the first Round-robin neural network parameters; Represents the gradient; Weighted hyperparameters are used to balance the data fitting loss and the physical consistency loss.

[0099] During the backpropagation phase, the network parameters are updated by the sum of the two gradients mentioned above. Simultaneously, the angle between the two gradient vectors and their magnitude ratio reflect the degree of cooperation or conflict between physical constraints and the data fitting objective. To construct a quantitative metric independent of absolute scale to measure this relationship, this invention normalizes the norm of the sum of the two gradients, defining a normalized synthetic gradient norm. as follows:

[0100] ;

[0101] This index takes values ​​in the range [0,1]. When the angle between the directions of two gradient vectors approaches 180 degrees, i.e., they are in opposite directions, and the absolute value of the difference between their magnitudes is less than a preset magnitude difference threshold, the index value approaches 0. The more consistent the directions of the two gradient vectors are, the closer its value is to 1.

[0102] In one round, there are a total of The training sample sequences are used for model training, therefore we use batch averaging to obtain the first set of training samples. Normalized composite gradient norm of rounds :

[0103] ;

[0104] Given the lag effect of errors in physical constraint equations on neural network parameter updates, this invention employs a sliding window averaging method to perform temporal smoothing on the normalized synthetic gradient norm for each training epoch, and uses the processed result as the final gradient conflict metric.

[0105] ;

[0106] In the formula: To adjust the sliding window size, For the first The normalized composite gradient norm after window averaging of rounds, i.e., the first round... The conflict metric for each round; v is the offset within the window; No. The normalized gradient norm of the rounds.

[0107] when When the gradient norm threshold is exceeded, the physical consistency loss stops participating in the model parameter update.

[0108] In summary, the robust physical information neural network prediction model constructed in this invention, by constraining the physical consistency between predicted and theoretical values ​​and supplementing it with a physical constraint exit mechanism based on conflict metrics, aims to maintain robust learning of the physical laws of the real system even when the theoretical model is imperfect or contains biases, thereby improving the model's robustness and physical consistency. The constructed net load prediction algorithm framework based on the robust physical information neural network is as follows: Figure 2 As shown.

[0109] S4: During model training, the normalized synthetic gradient norm is used as a conflict metric. When the conflict metric after sliding window averaging exceeds a preset threshold, the physical consistency loss stops participating in model parameter updates, and model training is driven only by the data fitting loss. Otherwise, the network parameters are updated by the weighted sum of the data fitting loss and the physical consistency loss. Through repeated iterative training, the net load of the desert area under extreme weather conditions of sandstorms is finally accurately predicted.

[0110] Specifically, to address the challenge of extracting time-series features from net load data caused by the strong fluctuations and high uncertainty of complex disturbances during sandstorms, this invention introduces the Mamba-SE prediction network to improve prediction performance in extreme weather scenarios.

[0111] The Mamba algorithm is a novel architecture for long sequence modeling. Based on the Selective Scan Structured State Space Sequence Model (S6), it replaces the self-attention mechanism in the traditional Transformer with a continuous-time state-space equation, achieving long-range dependency capture with linear complexity. Its key idea is that by introducing an input-dependent state-space parameter modulation mechanism, the model can adaptively adjust the weights of input and historical information in the state space, thereby accurately capturing key input information while automatically filtering out irrelevant noise and invalid long-term memories.

[0112] Specifically, at time step t, the S6 module receives the current input vector and the hidden state from the previous time step to update the hidden state at the current time step and form the output. Simultaneously, S6 uses a discretization transformation to map the continuous-time system dynamics equations into a recursion over discrete time steps. The hidden state update formula is:

[0113] ;

[0114] ;

[0115] In the formula: It is the hidden state at the current moment. It is the hidden state from the previous moment. , These are the input vector and the output vector, respectively. , and These correspond to the state transition, input projection, and readout discretization matrices, respectively. This represents the discretization factor related to the step size, which controls the mapping from a continuous-time system to a discrete-time system.

[0116] In prediction tasks, directly expanding the last time step of the Mamba algorithm to the desired output length might miss crucial dependency information carried by intermediate time steps, as complete dependency information is distributed across all time steps. Therefore, pooling is required for all time steps before reconstructing the output using a decoder to fully extract all dependencies in the sequence. This is the traditional prediction approach based on the Mamba algorithm.

[0117] However, traditional methods introduce global temporal feature information during pooling, which may contain not only useful information but also irrelevant or even interfering components. To alleviate this problem, this invention introduces a compression and activation channel attention mechanism before pooling and decoding to selectively emphasize and suppress features before aggregation, thereby improving the expression of effective dependencies and reducing the impact of useless or interfering information on the final prediction.

[0118] Meanwhile, the SE module recalibrates to highlight the feature channels that are more critical to the prediction results, allowing the constraint information in the loss function to be applied more concentratedly to these important channels, rather than being scattered across all feature channels. For the RPINN mechanism constructed earlier, this feature selection method helps to improve the efficiency of physical constraints in the feature learning stage.

[0119] Specifically, the prediction algorithm framework based on Mamba-SE is as follows: Figure 3 As shown in the figure, H represents height; This represents matrix multiplication. At this point, the SE module first outputs the Mamba algorithm... Global compression is performed on the feature dimensions to obtain the global statistics z for each channel, and then channel weights are generated through linear network activation. Finally, the temporal feature channels are recalibrated based on the obtained weights, thereby enhancing important channels and suppressing secondary channels. The process can be written as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula: and These represent the squeeze and excitation operations, respectively. The number of time-series characteristic channels; Indicates the first The global statistics obtained by one-dimensional global average pooling of each time-series feature channel ( (The eigenvector formed by all channel statistics). The time step length of the input feature sequence; For inclusion The set of learnable parameters, where, This represents the learnable parameter matrix in the fully connected layer. This represents the Sigmoid activation function; Represents the linear rectification activation function; To incentivize the output channel weight vector, its th... The elements are ; The feature tensor output by the Mamba algorithm represents the first feature tensor in the nth region. The time step, the first Feature values ​​on each feature channel ( (represents the complete time series on the k-th channel); This indicates that after recalibration by the Mamba-SE module, the final output is at the [missing information]. The time step, the first Feature values ​​on each channel.

[0124] Reference Figures 5-19 This is the second embodiment of the present invention, which provides a method for predicting net load in desert areas under sandstorm weather based on RPINN. In order to verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0125] Specifically, this embodiment uses load and new energy base operation data from extreme sandstorm weather in a typical desert region in Northwest my country over the past three years. The corresponding meteorological data are all from local meteorological stations, and the sampling frequency for all data is once per hour. The time range of the sandstorm event samples is selected from the atmospheric environment bulletin provided by the China Meteorological Administration's government portal website. This invention divides the training and test sets in an 8:2 ratio and uses Min-Max normalization on the input data. The normalization parameters are determined only by the minimum and maximum values ​​of the training set and applied to the test set to avoid data leakage.

[0126] In terms of scenario setup, this invention sets up two net load forecasting scenarios: 24-hour and 72-hour. The former is used to evaluate the short-term forecasting accuracy of the proposed method, while the latter is used to examine the proposed method's ability to characterize key features of long-term time-series inputs in forecasting tasks with longer time scales, thereby comprehensively verifying the generalization ability of the proposed method in forecasting tasks with different time scales.

[0127] Regarding evaluation indicators, for the accuracy of new energy power generation forecasting, this invention uses Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) for evaluation; for load and net load forecasting, MAE and Mean Absolute Percentage Error (MAPE) are used for quantification.

[0128] The correlation between photovoltaic power, wind power, load, and various meteorological data in the original data was analyzed using the Pearson correlation coefficient method, resulting in the correlation heatmap shown below. Figure 5 As shown.

[0129] As shown in the figure, photovoltaic power has the highest coupling degree with meteorological factors. The most correlated meteorological features are, in descending order, total ground irradiance, horizontal radiation, visibility, and normal radiation, with correlation coefficients of 0.997, 0.996, 0.882, and 0.878, respectively. Among them, visibility, as an important indicator for measuring the level of sandstorms, has a strong positive correlation with photovoltaic power, which significantly reflects the physical process of aerosols blocking radiation and thus inhibiting photovoltaic output during sandstorms.

[0130] Wind power output showed a strong correlation only with surface wind speed and meridional wind, with coefficients of 0.843 and -0.66, respectively. Wind speed is also a key indicator for measuring the intensity of sandstorms. This strong correlation indicates that wind power output is highly susceptible to sandstorm events. Furthermore, under extreme weather conditions during sandstorms, wind turbines may experience a sharp drop in power due to triggering shutdown protection.

[0131] Regarding load, the most relevant meteorological features, in descending order, are air temperature, total surface irradiance, humidity, and visibility, with correlation coefficients of 0.81, 0.66, -0.63, and 0.60, respectively. Although air temperature remains the primary factor influencing load changes, the impact of extreme weather events like sandstorms on load cannot be ignored. In summary, this invention will utilize the aforementioned meteorological data for experimental analysis.

[0132] To adapt to the indirect net load forecasting process, this embodiment inputs the meteorological variables selected through correlation analysis and each power series into TimeGan after constructing three sets of multivariate time-series datasets. By jointly inputting power and meteorological data, TimeGan can simultaneously capture the time-series autocorrelation of power and the coupling characteristics between data points, thereby maintaining the cross-variable correlation of the generated samples. Meanwhile, to prevent the generated data from contaminating the test set and thus affecting the evaluation results, TimeGan only generates and expands data on the training set; the generated samples are not included in the test set.

[0133] In this example, both the TSTR (Train on Synthetic, Test on Real) and TRTS (Train on Real, Test on Synthetic) metrics are measured using normalized mean squared error. The discriminant score is used to assess the similarity of the generated distributions; this metric is defined as the absolute value of the difference between the accuracy of an independent discriminator and 0.5. Table 1 shows the evaluation results of the generated data. Overall, the predictive scores of all three sets of data remain at a low level, demonstrating the usability of synthetic data in downstream prediction tasks. Simultaneously, the discriminant score and MMD metric are also small, indicating that the difference between the generated and real distributions is limited, and the fitting effect is good.

[0134] Table 1 Data Generation Results

[0135] .

[0136] Based on the correlation screening results, this embodiment constructs a generalized partial differential equation describing the power load response relationship using air temperature, total ground irradiance, humidity, visibility, and time coding. After solving the equation coefficients using PINN, a power load response relationship model with definite coefficients is obtained, and its residual histogram is shown below. Figure 6 As shown in the residual histogram, the load response model obtained by PINN has a good fitting ability to the data sample: the residual distribution is centered around 0 and exhibits a clear unimodal shape, with a mean of approximately 7.13 × 10⁻⁶. -4 This indicates that the overall systematic bias is relatively weak; at the same time, Approximately 7.13 × 10⁻⁶ -3 With -1.66×10 -3 The equation residuals for most samples are concentrated around 10. -3 Within the order of magnitude, this indicates that the model can effectively characterize the power load response pattern, providing a basis for using the residuals of this equation as the physical consistency loss for constructing the load forecasting sub-model.

[0137] Furthermore, the residual distribution exhibits a longer tail in the positive direction, with a small number of large positive residual points. Although these samples are not the main source of overall error, they reveal the asymmetric distribution characteristics of the fitting error on extreme samples, which may be related to short-term jumps or spikes in load under complex disturbances of dust storms.

[0138] To verify the adaptability of the proposed method under small sample conditions, this embodiment sets up four comparison methods to verify the adaptability of the proposed model under small sample conditions. For ease of discussion, the Mamba-SE model co-optimized by TimeGan and RPINN is denoted as Method 1, the Mamba-SE model based solely on the RPINN architecture is denoted as Method 2, the Mamba-SE model incorporating only TimeGan is denoted as Method 3, and the single Mamba-SE model is denoted as Method 4. In the above methods, the generated samples are not filtered using the MMD index. This example compares the prediction performance of the above methods and the method of this invention under four different original dataset proportions ranging from 25% to 100%. The experimental results are as follows: Figure 7 and Figure 8 As shown.

[0139] Experimental results show that in both time scales of net load forecasting tasks, the prediction accuracy of all methods decreases as the size of the original dataset decreases. When data is sufficient, Method 1 has the highest accuracy among the four comparison methods, followed by Method 2, then Method 3, and finally Method 4 has the lowest accuracy. However, when the proportion of original data is less than or equal to 50%, the accuracy of Method 4 is actually lower than that of Method 2.

[0140] In terms of data sensitivity, Method 4 was most significantly affected. Taking the 24-hour net load forecasting task as an example, when the original dataset was reduced from 100% to 25%, the prediction error increased from 2.582% to 6.969%, a 2.699-fold increase. Method 1 was the second most affected, with the prediction error increasing from 1.694% to 4.245%, a 250.6% increase. Method 3 was the third most affected (with an error increase of 200.1%). Method 2, on the other hand, was the most robust, with the prediction error increasing from 2.029% to 3.383%, a mere 1.667-fold increase.

[0141] When the proportion of original data is less than or equal to 50%, the accuracy of Method 1 is actually lower than that of Method 2, which is less sensitive to data volume. This is because, although both TimeGan and RPINN can alleviate data dependency when used individually, their direct combination under low data volume can lead to negative coupling. When original data samples are scarce, the quality of data generated by TimeGan decreases significantly. This not only directly reduces the input quality of the neural network but also severely reduces the accuracy of the meteorological data upon which the RPINN mechanism relies. Even though the RPINN mechanism has introduced a physical constraint exit mechanism based on conflict metrics, low-quality meteorological data will still cause physical constraints to have a negative impact on model training. This dual negative effect makes the collaborative model more sensitive to data reduction than when each component operates independently.

[0142] However, the method of this invention achieved the lowest error under different original dataset sizes. This is due to the adaptive adjustment capability given to the model by the MMD screening mechanism: when the original data is sufficient, high-quality generated data is included in the original dataset, and the method of this invention is equivalent to Method 1, thereby obtaining the highest accuracy; while when the original data is scarce, inferior generated samples are eliminated, and the method of this invention becomes equivalent to Method 2. By eliminating low-quality generated data, it achieves the best performance in small sample scenarios.

[0143] To verify the necessity of the robust physical information neural network mechanism proposed in this invention, this embodiment compares the training process of the Mamba-SE model under three modes: RPINN, traditional PINN, and no PINN. Figures 9-11 As shown in the figure. It should be noted that, since the conflict metric in photovoltaic forecasting is always higher than this threshold, the figure only shows the training process of the load and wind power forecasting models that triggered the physical constraint exit mechanism.

[0144] Experimental results show that, without the physical constraint exit mechanism, although the model's prediction error is still lower than without PINN, a significant "reverse increase" phenomenon occurs during training: the error starts to rise again after decreasing to a certain point and increases significantly with the number of training rounds. By introducing a conflict-based physical constraint exit mechanism, when the conflict between the data fitting loss and the physical consistency loss becomes too large, the model adaptively pauses the physical consistency loss from participating in parameter updates. This allows the prediction error to continue decreasing even after it has reached its minimum value, resulting in better prediction performance.

[0145] It is worth noting that even though the conflict metric during the model training stabilization period is significantly improved when using the traditional PINN method, the use of a robust physical information neural network mechanism is still necessary. This is because the intense conflict between the two loss terms in the early stages, although only lasting for a short period, can still significantly mislead the model, leading to poor final results. Deep learning training is a complex nonlinear dynamic process and is extremely sensitive to initial conditions. If intense gradient conflicts in the early stages are not masked in time, they can significantly and persistently mislead the model, even causing the model parameters to get stuck in local optima. This bias is difficult to self-correct through later training; instead, it accumulates and spreads in the network with each training iteration, preventing the model from consistently learning on the correct track and severely limiting the final prediction performance.

[0146] To evaluate the predictive performance of the proposed RPINN-Mamba-SE model in dust storm scenarios, this embodiment selects three mainstream models—Mamba, Informer, and Bi-LSTM (Bidirectional Long Short-Term Memory)—and conducts comparative experiments on net load prediction at two time scales. The prediction results are shown in […]. Figures 12-15 and Figures 16-19 The specific prediction errors are shown in Tables 2 and 3. Overall, the RPINN-Mamba-SE model proposed in this invention achieves the best prediction accuracy at both time scales, with net load prediction errors of 1.694% and 1.888%, respectively, which are significantly better than the aforementioned mainstream models.

[0147] To further verify the contributions of the RPINN mechanism and the SE module, this invention designs an ablation experiment around the complete RPINN-Mamba-SE model to examine the changes in the predictive performance of the model after the two modules are gradually ablated. Experimental results show that the robust physical information mechanism and the SE module have significant complementary gains. On the one hand, physical constraints enhance the effectiveness of the SE module: in both 24-hour and 72-hour predictions, the performance improvement of RPINN-Mamba-SE compared to RPINN-Mamba is superior to that of Mamba-SE compared to Mamba. Specifically, taking the 24-hour net load prediction results as an example, the former optimizes the MAE from 78.77MW to 56.42MW, an improvement of 28.37%, higher than the latter's 19.40% improvement. This is because the physical consistency loss, by constraining the gradient direction and suppressing noise caused by complex disturbances in dusty weather, makes the learning of SE channel weights more stable and efficient. On the other hand, the SE module makes the introduction of physical constraints more effective: the improvement of RPINN-Mamba-SE over Mamba-SE (from 80.16MW to 56.42MW) is significantly higher than the improvement of RPINN-Mamba over Mamba (from 99.46MW to 78.77MW). This is because the SE module selects key feature channels through recalibration, making it easier for physical constraints to focus on key channels that are more important to the prediction results, thereby giving fuller play to the constraint role of physical laws and improving the efficiency of physical constraints.

[0148] Table 2 Comparison of 24-hour forecast results .

[0149] Table 3 Comparison of 72-hour forecast results .

[0150] As demonstrated by the above embodiments, this invention proposes a net load forecasting method for desert and Gobi regions under sandstorm weather based on RPINN. This method addresses the shortcomings of existing research in meeting the accurate net load forecasting requirements for power grid optimization and dispatching in desert and Gobi regions under sandstorm weather, as well as the limitations of data acquisition conditions leading to a scarcity of training samples and difficulties in extracting key time-series features due to complex disturbances in sandstorm scenarios. To meet the accurate net load forecasting needs of power grid optimization and dispatching in desert and Gobi regions under extreme sandstorm weather, this invention constructs a targeted net load forecasting framework to enhance its support for dispatching decisions. It introduces generalized partial differential equations to construct a load forecasting model with embedded physical constraints, and builds a robust physical information neural network based on traditional PINN to enhance the model's learning robustness under small sample conditions. A Mamba-SE algorithm incorporating the SE mechanism is proposed to improve the efficiency of physical constraint action while extracting key time-series feature information from highly volatile and uncertain inputs under complex disturbances. Experimental results show that the proposed method can effectively alleviate the above problems, thus achieving better prediction performance.

Claims

1. A method for predicting net load in desert areas under sandstorm weather based on RPINN, characterized in that, Includes the following steps: Step 1: To address the lack of historical samples under extreme weather conditions such as sandstorms, Pearson correlation coefficient is used to analyze the correlation between photovoltaic, wind power, and load data and meteorological data under extreme weather conditions such as sandstorms. The obtained highly correlated data sets are input into the time-based generative adversarial network model for data augmentation. The generated data with the maximum mean difference less than a set threshold are selected and combined with the original data to form an enhanced dataset, which is used as the input for the subsequent prediction network. Step 2: Based on the correlation analysis results of the load data and meteorological data obtained in Step 1, a generalized partial differential equation for load response relationship modeling is constructed. The coefficients of the generalized partial differential equation are adaptively solved using the Physical Information Neural Network (PINN) to obtain a power load response relationship model with definite coefficients, so as to construct the physical constraints for load forecasting under sandstorm weather. Among them, the difference between the partial derivatives of power with respect to key environmental factors is used as the physical consistency loss. Step 3: After expanding the feature dimension of the input data through the encoding layer, the feature sequence is encoded and then enters the Mamba-SE module. After pooling and decoding operations, the final prediction result is output. The Mamba-SE module introduces a compression and activation channel attention mechanism before the pooling and decoding of the Mamba algorithm to selectively recalibrate the temporal feature channels. Step 4: During model training, the normalized synthetic gradient norm is used as the conflict metric. When the conflict metric after sliding window averaging exceeds the preset threshold, the physical consistency loss is stopped from participating in the model parameter update, and the model training is driven only by the data fitting loss; otherwise, the network parameters are updated by the weighted sum of the data fitting loss and the physical consistency loss. Through repeated iterative training, the net load of the desert area under the extreme weather conditions of sandstorms is finally accurately predicted.

2. The method for predicting net load in desert areas under sandstorm weather based on RPINN according to claim 1, characterized in that, In step 1, the formula for calculating the Pearson correlation coefficient is as follows: ； In the formula, , The first Each observed object in the variable , The value of ; , The mean of the sample; Indicates the sample length of the time series; For variables , The Pearson correlation coefficient; After feature selection, TimeGan is introduced to augment the data. TimeGan combines adversarial learning of GAN with temporal modeling and introduces supervision constraints based on latent space, so that the generated sequence is consistent with the real data in terms of statistical distribution while maintaining the temporal dependency structure. The maximum mean difference is used to measure the distribution consistency between generated data and real data. Generated data is included in the augmented dataset only when the maximum mean difference between generated data and real data is less than a pre-set threshold. Let the real sample set be The generated sample set is In kernel function Below, the maximum mean difference is: ； In the formula, The maximum mean difference between the real sample set X and the generated sample set Y; and These are the sample sizes of the real sample set and the generated sample set, respectively. Take the Gaussian radial basis function kernel; and These are real samples; and To generate samples.

3. The method for predicting net load in desert areas under sandstorm weather based on RPINN according to claim 1, characterized in that, In step 2, the indirect method is used to predict the system net load: three independent neural network sub-models are constructed for photovoltaic power generation prediction, wind power generation prediction and power load prediction respectively, and the difference between the load prediction result and the photovoltaic and wind power prediction results is used as the system net load prediction result. The photovoltaic generator set adopts maximum power point tracking technology, and the photovoltaic power generation capacity is [missing information]. The calculation formula is as follows: ； In the formula, , These represent the number of solar cell strings and the number of parallel connections in the photovoltaic unit. and These are the maximum operating point current and maximum operating point voltage of the solar cell, respectively. Maximum operating point current under standard operating conditions With maximum operating point voltage With the introduction of a compensation coefficient, the corrected formulas for the operating point current and operating point voltage are as follows: ； ； ； ； In the formula, , These represent irradiance and temperature, respectively. , These are standard irradiation intensity and standard temperature, respectively. , and This is an empirical coefficient; The corrected operating point current, This is the corrected operating point voltage; , These are the deviations between the actual temperature and the standard temperature, and the relative deviations between the actual irradiance and the standard irradiance, respectively. It is a natural constant; The first derivative of photovoltaic power generation with respect to solar irradiance can be calculated using the following formula: ； Power generation of a single wind turbine under extreme weather conditions of a sandstorm Wind turbine generator shutdown state control function and wind turbine generator power generation They are respectively: ； ； ； In the formula, The power generation capacity of the wind turbine generator set, The power generation capacity of a single wind turbine. air density, For wind speed, The radius of the swept surface of the wind turbine; Power coefficient; This refers to the number of wind turbines within the wind turbine generator set. Indicates the time when the sandstorm occurred. This indicates the delay in wind farm shutdown after the disaster; , and These are the cut-in wind speed, cut-out wind speed, and rated wind speed of the wind turbine generator set, respectively. The rated power output of a single wind turbine. This is a control function for the shutdown status of wind turbine generators under extreme weather conditions such as sandstorms. A value of 1 indicates normal operation, and a value of 0 indicates shutdown due to disaster. The differential symbol; The first derivative of wind turbine generator power with respect to wind speed is calculated using the following formula: ； ； The general mathematical form of the generalized partial differential equation is: ； in, The coefficients of the generalized partial differential equation; The unknown function being fitted represents the system state or output; , where is the independent variable or the system's input variable; D is the number of independent variables affecting load changes; unknown function To represent the power system load, input variables will be used. It represents the various factors affecting load changes, including time-coded and meteorological factors screened by correlation; it uses a physical information neural network to adaptively solve the equation coefficients to obtain a load response relationship equation with definite coefficients, which is used to characterize the dynamic response law of load to multiple factors.

4. The method for predicting net load in desert areas under sandstorm weather based on RPINN according to claim 3, characterized in that, In step 3, the Mamba-SE module specifically includes: The Mamba algorithm is based on a selective state-space model S6. At time step t, the selective state-space model S6 receives the current input vector. and the hidden state of the previous moment Used to update the hidden state at the current time step and form the output vector The discretization update formula is: ； ； in, It is in a hidden state; , These are the input vector and the output vector, respectively. , and These correspond to the state transition, input projection, and readout discretization matrices, respectively. This represents the discretization factor related to the step size; The compression and excitation channel attention mechanism is described in the Mamba algorithm output. First, global compression is performed on the feature dimension to obtain the global statistics z for each channel. Then, channel weights s are generated through linear network activation. Finally, the temporal feature channels are recalibrated based on the obtained weights to enhance important channels and suppress minor channels. The process is as follows: ； ； ； In the formula, and These represent compression and excitation operations, respectively. The number of time-series characteristic channels; Indicates the first The global statistics obtained by one-dimensional global average pooling of each time-series feature channel; The feature vector formed by all channel statistics; The time step length of the input feature sequence; For inclusion The set of learnable parameters, where, This represents the learnable parameter matrix in the fully connected layer. This represents the Sigmoid activation function; Represents the linear rectification activation function; To incentivize the output channel weight vector, its th... The elements are ; The feature tensor output by the Mamba algorithm represents the first feature tensor in the nth region. The time step, the first Feature values ​​on each feature channel; This represents the complete time series on the k-th channel; This indicates that after recalibration by the Mamba-SE module, the final output is at the [missing information]. The time step, the first Feature values ​​on each channel.

5. The method for predicting net load in desert areas under sandstorm weather based on RPINN according to claim 4, characterized in that, In step 4, the normalized synthesis gradient norm is calculated as follows: The loss functions for the three neural network sub-models are as follows: ； ； ； In the formula, , and These represent the total losses of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and electricity load forecasting, respectively. , and These represent the data fitting losses for photovoltaic power generation forecasts, wind power generation forecasts, and electricity load forecasts, respectively. , and The physical consistency loss of photovoltaic power generation forecasting, wind power generation forecasting and electricity load forecasting; , and These represent the weighting coefficients of the data fitting loss and physical consistency loss of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and power load forecasting, respectively. The data fitting loss for the three neural network sub-models is calculated using the mean squared error between the predicted and actual values, as follows: ； ； ； In the formula, , and These represent the data fitting losses of the three sub-models for photovoltaic power generation forecasting, wind power generation forecasting, and electricity load forecasting, respectively; M is the total number of sample sequences used in training the sub-models. , , The first generation of the three neural network sub-models The true output power value of each training sample sequence; , , The first generation of the three neural network sub-models Predicted output power values ​​for each training sample sequence; Physical consistency loss of photovoltaic and wind power prediction models and They are respectively: ； ； In the formula, and These are the measured irradiance and measured wind speed, respectively, from the model database; and These are the theoretical outputs of photovoltaic and wind power, calculated from measured data, respectively. , These are the output power prediction values ​​generated by the neural network sub-models for photovoltaic power generation prediction and wind power generation prediction, respectively. Physical consistency loss of load forecasting models To find the squared value of the residuals obtained by substituting the load response relationship equation with determined coefficients into the predicted and measured data: ； In the formula, This represents the residual value of the equation calculated by substituting the measured data and predicted data at the corresponding time into the generalized partial differential equation; Measured data for various factors affecting load changes; For each prediction model, at the 1st Rounds, recorded Loss on data fitting of training sample sequences loss of physical consistency The gradient is and : ； ； In the formula, For the first Round-robin neural network parameters; Represents the gradient; The weighting hyperparameters are used to balance the data fitting loss and the physical consistency loss. Define the normalized synthesis gradient norm as follows: ； The chemical gradient norm The value takes values ​​in the range [0,1]. When the angle between the directions of the two gradient vectors approaches 180 degrees, that is, when the directions are opposite, and the absolute value of the difference between their magnitudes is less than the preset magnitude difference threshold, the norm value approaches 0; the more consistent the directions of the two, the closer its value is to 1. No. Normalized composite gradient norm of rounds : ； A smoothed gradient conflict metric is obtained by averaging the gradients over each training round using a sliding window method. ； In the formula, To adjust the sliding window size, For the first The normalized composite gradient norm after window averaging of rounds, i.e., the first round... The conflict metric for each round; v is the offset within the window; No. The normalized composite gradient norm of the rounds; when When the gradient norm threshold is exceeded, the physical consistency loss stops participating in the model parameter update.

Images