Photovoltaic power station day-ahead generation power prediction method fusing numerical weather forecast information and scene perception strategy

Abstract

The invention discloses a photovoltaic power station day-ahead generation power prediction method fusing numerical weather forecast information and a scene perception strategy, and belongs to the field of photovoltaic generation power prediction. The method comprises the following steps: preprocessing historical power data and numerical weather forecast data of a photovoltaic power station; pre-processed photovoltaic power data and meteorological variable data are obtained; based on the preprocessed photovoltaic power data and meteorological variable data, analyzing the correlation between each meteorological variable and the photovoltaic power by adopting a Pearson correlation coefficient, and screening out key meteorological factors; inputting the key meteorological factors into a Gaussian mixture model for unsupervised clustering, and constructing a typical weather scene sample set; based on the typical weather scene sample set, training a corresponding convolutional residual memory network prediction model for each typical weather scene to obtain a plurality of sub-models; and fusing the plurality of sub-models and outputting a day-ahead power generation power prediction result of the photovoltaic power station.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation prediction technology, and in particular relates to a method for predicting the day-ahead power generation of photovoltaic power plants by integrating numerical weather forecast information and scene perception strategy. Background Technology

[0002] As a crucial component of clean energy, photovoltaic (PV) power generation has seen continuous growth in installed capacity globally. However, PV output exhibits significant intermittency and volatility, primarily due to its strong dependence on meteorological conditions for power generation efficiency. Dynamic changes in meteorological factors such as solar irradiance, cloud cover, and temperature lead to nonlinear and non-stationary characteristics in PV power output, posing a severe challenge to power balance and dispatching operations in the power grid. To improve prediction accuracy, a common approach is to combine numerical weather prediction data with historical power data, and to incorporate various machine learning models for modeling and analysis.

[0003] While existing technologies have improved prediction performance to some extent, significant limitations remain. On one hand, most methods do not fully consider the differences in the relationship between meteorological characteristics and power output under different weather types, resulting in insufficient adaptability of the models in complex weather transition scenarios. On the other hand, traditional prediction models often use a uniform structure to handle all weather conditions, making it difficult to accurately capture dynamic features under specific weather patterns, especially given the significant differences in prediction accuracy between typical scenarios such as sunny, cloudy, rainy, and special weather conditions. Furthermore, conventional deep learning models are prone to the vanishing gradient problem when dealing with long-term temporal dependencies and have limited ability to extract spatial features, hindering further improvements in prediction performance.

[0004] Therefore, this invention provides a method for predicting the day-ahead power generation of photovoltaic power plants by integrating numerical weather forecast information and scene perception strategies. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for predicting the day-ahead power generation of photovoltaic power plants by integrating numerical weather forecast information with scene perception strategies, thereby resolving the issues existing in the prior art.

[0006] To achieve the above objectives, this invention provides a method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather prediction information and scene perception strategies, comprising:

[0007] Historical power data and numerical weather forecast data of photovoltaic power plants are preprocessed to obtain preprocessed photovoltaic power data and meteorological variable data.

[0008] Based on the preprocessed photovoltaic power data and meteorological variable data, the correlation between each meteorological variable and photovoltaic power was analyzed using the Pearson correlation coefficient to screen out key meteorological factors.

[0009] The key meteorological factors are input into a Gaussian mixture model for unsupervised clustering to construct a sample set of typical weather scenarios.

[0010] Based on the typical weather scenario sample set, for each typical weather scenario, a corresponding convolutional residual memory network prediction model is trained to obtain several sub-models.

[0011] The aforementioned sub-models are fused together to output the day-ahead power generation prediction results for photovoltaic power plants.

[0012] Optionally, the key meteorological factors include: air temperature, total radiation, and cloud cover.

[0013] Optionally, the probability density function of the Gaussian mixture model is expressed as:

[0014] ;

[0015] In the formula, It is the true distribution of the whole. The estimate, Let be the weighting coefficients, mean, and variance of the k-th Gaussian component, respectively. Let be the k-th Gaussian component of the GMM.

[0016] Optionally, the convolutional residual memory network prediction model includes: a convolutional neural network, a residual connection mechanism, and a gated recurrent unit;

[0017] The process of outputting the power prediction result based on the convolutional residual memory network prediction model includes:

[0018] A convolutional neural network is used to extract spatial features from the input meteorological feature sequence to obtain high-dimensional spatial features;

[0019] The high-dimensional spatial features are fused using a residual connection mechanism to obtain an enhanced feature sequence.

[0020] A gated recurrent unit is used to model the temporal dependency relationship of the enhanced feature sequence to obtain a temporal feature representation;

[0021] The photovoltaic power generation prediction results are obtained by performing regression processing on the time-series feature representation based on the fully connected layer.

[0022] Optionally, the hidden layer output expression of the gated loop unit is:

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] In the formula, r t To reset the door, z t To update the gate, h t Currently in a hidden state. It is the Sigmoid activation function. It is the hyperbolic tangent function. For the current input, The state was hidden in the previous moment. In the candidate hidden state, and These represent resetting and updating gate weights, respectively; * represents the Hadamard product. The weight coefficients represent the candidate hidden states.

[0028] Optionally, the process of obtaining several sub-models includes: testing the corresponding training convolutional residual memory network prediction model with several evaluation metrics to obtain the optimal sub-model;

[0029] The evaluation indicators include: root mean square error, mean absolute error, mean error, coefficient of determination, pass rate, and accuracy.

[0030] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0031] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] This invention proposes a photovoltaic power prediction method that integrates historical power data and numerical weather forecast information. It combines unsupervised partitioning of typical weather scenarios with a scenario-specific modeling strategy of the CRMN model, effectively improving the model's ability to identify and adapt to complex meteorological changes. By training each type of weather scenario separately, the model's generalization performance under various operating conditions is enhanced. Simultaneously, the introduction of convolutional and residual memory structures improves the extraction capability and training efficiency of meteorological time-series features, avoiding the gradient vanishing problem common in traditional deep networks. The overall method possesses advantages such as high prediction accuracy, strong stability, fast response, and wide applicability, providing reliable data support and technical assurance for grid dispatch optimization and high-proportion renewable energy integration. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0035] Figure 1 This is a flowchart of a photovoltaic power generation prediction method that integrates numerical weather forecast information and scene perception strategy according to an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the CRMN model according to an embodiment of the present invention;

[0037] Figure 3 This is a diagram of the internal structure of the GRU according to an embodiment of the present invention;

[0038] Figure 4 The following are the relationships between power generation and major weather factors in embodiments of the present invention, wherein (a) is the relationship between power generation and wind speed, (b) is the relationship between power generation and air pressure, (c) is the relationship between power generation and solar radiation intensity, (d) is the relationship between power generation and humidity, (e) is the relationship between power generation and temperature, and (f) is the relationship between power generation and cloud cover.

[0039] Figure 5 The following diagrams show the clustering results obtained by the Gaussian mixture model in an embodiment of the present invention: (a) The covariance matrix is ​​a diagonal matrix and all clusters share the same covariance matrix; (b) The covariance matrix is ​​a diagonal matrix and all clusters have independent covariance matrices; (c) The covariance matrix is ​​a complete matrix and all clusters share the same covariance matrix; and (d) The covariance matrix is ​​a complete matrix and all clusters have independent covariance matrices.

[0040] Figure 6 The above are similarity clustering results obtained based on Gaussian mixture model in an embodiment of the present invention; wherein, (a) is the first type of sunny similar days, (b) is the second type of special weather days, (c) is the third type of cloudy similar days, and (d) is the fourth type of rainy similar days.

[0041] Figure 7 The present invention provides a power prediction under the segmented scenario, wherein (a) is the power prediction for the second month, (b) is the power prediction for the fifth month, (c) is the power prediction for the eighth month, and (d) is the power prediction for the eleventh month.

[0042] Figure 8 This is an evaluation of the model prediction accuracy in scenarios with and without scenario division, according to an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] Example 1

[0046] like Figure 1 As shown in the figure, this embodiment provides a method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy, including the following steps:

[0047] S1. Preprocess the historical power data and numerical weather prediction (NWP) data of photovoltaic power plants, including outlier removal, missing data completion and normalization operations.

[0048] S2. By analyzing the relationship between various meteorological factors and photovoltaic power through Pearson correlation coefficient analysis, key variables with high correlation such as temperature, total radiation, and cloud cover were selected.

[0049] S3. Cluster analysis of NWP variables is performed using Gaussian mixture models to automatically classify weather conditions, thereby constructing a sample dataset covering multiple typical weather categories;

[0050] S4. Train CRMN models separately for different weather scenarios to improve prediction accuracy under specific meteorological conditions;

[0051] S5. The performance of each sub-model is tested through multi-dimensional evaluation indicators, including root mean square error (RMSE), mean absolute error (MAE), mean error (ME), coefficient of determination (R²), pass rate (Q), and accuracy (Cr), and the optimal model is selected and integrated for the final prediction.

[0052] As a specific implementation method of this embodiment, the following model is specifically included:

[0053] Gaussian Mixture Model (GMM) is a soft clustering algorithm based on probability statistics. It assumes that the data is a mixture of multiple Gaussian distributions (i.e., normal distributions). Unlike traditional hard clustering methods (such as K-means), GMM does not uniquely assign each sample point to a specific class. Instead, it calculates the probability of a sample belonging to each Gaussian component by maximizing the posterior probability of the sample. In other words, a data point can belong to multiple classes simultaneously, but with different probabilities. This "soft assignment" mechanism better reflects the ambiguity and continuity of data in reality, making it suitable for processing datasets with overlapping boundaries or irregular distributions. GMM typically uses the Expectation-Maximization (EM) algorithm to estimate model parameters and has wide applications in clustering, density estimation, and anomaly detection.

[0054] Assume a sample point The probability density function of a GMM is defined as:

[0055]

[0056] In the formula: It is the true distribution of the whole. The estimate, Let be the weighting coefficients, mean, and variance of the k-th Gaussian component, respectively. The expression for the k-th Gaussian component of the GMM is as follows:

[0057]

[0058] GMM can model cluster structures of arbitrary shapes and overlaps, providing richer information than K-means through probability output. It can also be extended to complex problems such as high-dimensional data, weighted samples, and multimodal density estimation.

[0059] CRMN (Convolutional Residual Memory Network): Traditional Convolutional Residual Memory Networks (CRMNs) generally rely on Long Short-Term Memory (LSTM) networks as memory modules. This paper innovatively replaces LSTM with Gated Recurrent Units (GRUs), which significantly improves training efficiency while maintaining the model's expressive power and further enhances prediction performance.

[0060] like Figure 2As shown, CRMN consists of two layers of residual blocks and two layers of GRU units. The residual blocks focus on capturing the spatial features of the input data, which is key to the model's extraction of local structural information; while the GRU units focus on the temporal dynamics of the feature sequence, accurately capturing the patterns of change. The input first passes through the residual blocks, which generate a more discriminative high-dimensional spatial feature sequence through skip connections and convolution operations. This sequence serves as the input to the GRU. It's worth noting that the GRU here does not process traditional time series, but rather sequences based on spatial feature dimensions, where the time dimension is replaced by the sequential feature dimension. Finally, the outputs of the residual blocks and the GRU are fused point-by-point to form a unified feature vector, which is then passed to the regression layer to accurately predict photovoltaic power generation.

[0061] The Gated Recurrent Unit (GRU) is a simplified version of the Long Short-Term Memory (LSTM) network, belonging to the same category of Recurrent Neural Networks (RNNs). Its core structure consists of two gating mechanisms: a reset gate and a recurrent gate. t and update gate z t The activation functions are Sigmoid and hyperbolic tangent (Tanh), respectively. The reset gate regulates the degree of forgetting, focusing on the filtering and resetting of short-term information; the update gate controls the proportion of historical information retained by the hidden state, ensuring the effective transmission of long-term dependencies. Let the candidate hidden state be h, and the current hidden state be h. t The input is x, and element-wise multiplication is represented by symbols. The final hidden state is generated by weighted summation ⊕. This design effectively alleviates the gradient vanishing problem and improves the stability and efficiency of sequence modeling.

[0062] like Figure 3 As shown in the diagram, the calculation formula for the hidden layer output can be derived from the internal structure diagram of GRU as follows:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] The features and objective function are described as follows:

[0069] (1) Meteorological characteristics

[0070] The photovoltaic power generation prediction method of this invention uses numerical weather prediction (NWP) information as the key input feature of the model to fully explore the influence of meteorological conditions on photovoltaic power generation. Specifically, the NWP information used includes multi-dimensional meteorological parameters such as wind speed, air pressure, solar irradiance, humidity, temperature, and cloud cover. These parameters are all from authoritative numerical weather prediction systems and have high spatiotemporal resolution and strong prediction accuracy.

[0071] Wind speed, as one of the meteorological variables, affects the operating environment and cooling effect of photovoltaic power station equipment, indirectly impacting power generation efficiency; air pressure reflects changes in atmospheric pressure, is related to the evolution of weather systems, and affects cloud formation and atmospheric transparency; solar irradiance directly determines the light energy input of photovoltaic modules and is the most critical meteorological variable affecting photovoltaic power generation; humidity reflects the water vapor content in the air, affecting atmospheric transmittance and the surface condition of photovoltaic modules, thus affecting power generation performance; temperature is an important factor affecting the conversion efficiency of photovoltaic modules, and excessively high or low temperatures can lead to fluctuations in power generation; cloud cover affects illumination conditions, and cloud thickness and distribution directly relate to fluctuations in illumination intensity, reflecting the randomness and uncertainty of photovoltaic power generation.

[0072] The Pearson correlation coefficient method was used to calculate the correlation between various meteorological variables (such as wind speed, air pressure, solar irradiance, humidity, temperature, and cloud cover) and power generation, and variables were selected accordingly. Figure 4 As shown, when the correlation coefficient between a variable and power is significantly low (e.g., below a preset threshold), it indicates a weak linear explanatory power for power generation, potentially contributing little to the model's predictive performance and even leading to overfitting or redundant features. Therefore, this invention, based on correlation ranking results, retains meteorological variables strongly correlated with power generation as the final input features, constructing a concise yet representative input feature set. This feature selection strategy effectively reduces the input dimensionality while ensuring the model retains key meteorological information, thus improving the model's training efficiency and generalization ability.

[0073] (2) Design of CRMN photovoltaic power prediction model

[0074] Replacing the Long Short-Term Memory (LSTM) network in a traditional Convolutional Residual Memory (CRMN) network with a Gated Recurrent Unit (GRU) not only simplifies the structure but also significantly improves the model's computational efficiency and training speed. By reducing the number of gating units and the size of parameters, GRU reduces model complexity without sacrificing expressive power, facilitating rapid training and deployment under limited computing power. Furthermore, GRU possesses excellent temporal modeling capabilities, effectively capturing dynamic trends in input feature sequences. Compared to LSTM, it demonstrates greater stability and adaptability when handling short- to medium-term dependencies.

[0075] In photovoltaic power generation prediction tasks, input features exhibit significant temporal and fluctuating characteristics. GRU is more sensitive to the evolution of key meteorological features over time, contributing to improved prediction accuracy. Furthermore, its simplified structure mitigates overfitting risks, allowing the model to maintain good generalization ability even with limited training data or significant data volatility. Therefore, replacing LSTM with GRU not only enhances the model's practicality and stability but also provides a more feasible technical path for efficient photovoltaic power generation prediction.

[0076] (3) Model evaluation

[0077] To comprehensively evaluate the performance of the photovoltaic power generation prediction model constructed in this paper, several typical regression prediction models will be built, including traditional time series models (such as ARIMA), machine learning models (such as random forest RF), and deep learning models (such as long short-term memory network LSTM, convolutional neural network CNN, etc.), and compared and validated at both quantitative and visual analysis levels. On the one hand, by introducing various error evaluation indicators, including root mean square error (RMSE), mean absolute error (MAE), mean error (ME), coefficient of determination (R2), accuracy (Cr), and pass rate (Q), the difference between the model prediction results and the actual observed values ​​is quantified from different perspectives.

[0078] For embodiments of the present invention, simulation experiments were conducted using power generation and meteorological data from a photovoltaic power station in Uluru, Australia, in 2018 as samples. Data from the last week of months 2, 5, 8, and 11 were used as the test set, while other data served as the training set. The prediction time range was 7 days, and the time resolution was 15 minutes. The Pearson correlation coefficients between various meteorological conditions and power were calculated and are shown in Table 1, revealing a significant correlation between photovoltaic power and different meteorological characteristics. Specifically, the correlation coefficient between temperature and power was 0.3649, and that between irradiance and power was 0.8628, indicating that stronger temperature and irradiance resulted in higher module output power. The correlation coefficient between cloud cover and power was -0.327, showing a negative correlation: when there is ample sunshine and little cloud cover, photovoltaic power is less affected; however, when there is abundant cloud cover, increased shading of solar radiation significantly reduces photovoltaic output.

[0079] Table 1

[0080] feature Correlation coefficient Temperature 0.3649 radiation 0.8628 Cloud cover -0.327

[0081] Without defining meteorological scenarios, the evaluation indicators of the six models were compared, and the results are shown in Table 2. As can be seen from the table, the CRMN model performed best in all evaluation indicators, with an accuracy (Cr) of 97.0436% and a pass rate (Q) of 99.9628%. The experimental results indicate that by incorporating NWP information, the model can accurately capture the impact of NWP information on photovoltaic power output, thereby improving its predictive performance.

[0082] Table 2

[0083] Model RMSE MAE ME R2 Cr Q CRMN 0.0296 0.0166 0.0006 0.9893 97.0436 99.9628 LSTM 0.0373 0.0185 0.0007 0.983 96.2734 99.8512 CNN 0.0378 0.0188 0.0015 0.9824 96.2151 99.8884 1D-CNN 0.0402 0.0197 0.0009 0.9802 95.9756 99.7768 RF 0.0952 0.0564 0.0264 0.889 90.4826 98.0655 ARIMA 0.1220 0.0726 0.0519 0.8178 87.8041 90.3274

[0084] Unsupervised clustering of meteorological conditions throughout the year was performed using a Gaussian mixture model, resulting in four clustering outcomes for similar days, such as... Figure 5 As shown in the figure, each graph represents a different assumption about the covariance structure, under which the data is divided into four different clusters (represented by different colors in the graph), with "X" markings indicating cluster centers. In the top left graph, all clusters share the same covariance matrix, resulting in four relatively clear clusters. While the clusters are relatively independent, their expressive power is weak. In the top right graph, the shape and distribution of the clusters differ from the top left graph, reflecting that different clusters can have their own feature distributions under this result, but it cannot accurately represent the correlation between variables. In the bottom left graph, although the correlation between variables can be expressed, the cluster shape is relatively uniform, limiting its expressive power. In the bottom right graph, correlation between features is allowed within each cluster, and each cluster has a unique feature correlation structure. The shape and distribution of the clusters are more diverse, better reflecting the true structure of the data. To more accurately predict photovoltaic power, the clustering results need to be more precise. Considering the potential correlation between meteorological conditions, the bottom right graph was chosen as the clustering result for this study.

[0085] Figure 6 The results of similarity clustering obtained based on the Gaussian mixture model in this embodiment of the invention have been normalized to per unit. In the original cluster count, the cluster center curve of the first scenario shows a clear unimodal distribution, with a significant peak at midday. Therefore, the first scenario is classified as a sunny-day similarity. The second scenario also shows a unimodal distribution, with its power generation cluster curve significantly smaller than that of the first, third, and fourth scenarios. Furthermore, the power generation exhibits significant fluctuations and randomness, possibly due to special weather conditions such as typhoons, heavy rain, and blizzards. Therefore, the second scenario is classified as a special weather day. The power generation difference between the third and fourth scenarios is approximately 10kW, so the third scenario is classified as a cloudy-day similarity and a rainy-day similarity. The results are shown in Table 3.

[0086] Table 3

[0087] Clustering results First type of scenario Second type of scenario Third type of scenario Fourth type of scenario Scene summary Sunny Day Special Weather Day Cloudy day Rainy Day Similar to Day

[0088] The first scenario has 65 days, the second scenario has 63 days, the third scenario has 151 days, and the fourth scenario has 86 days. Based on the clustering results, February and August belong to the fourth scenario; May belongs to the first scenario; and November belongs to the second scenario. The CRMN model was trained on each of the four clustering scenarios to obtain CRMN prediction models for different scenarios. Power predictions were then performed for February, May, August, and November, and the prediction results are as follows: Figure 7 As shown.

[0089] The evaluation metrics of the CRMN models with and without scene partitioning are plotted in radar charts as follows. Figure 8 As can be seen, all metrics under scenario segmentation are superior to those without. Specifically, scenario segmentation reduces RMSE by 14.86%, MAE by 9.63%, R² by 0.29%, Cr by 0.45%, and Q by 0.04%, achieving a 100% pass rate. This is because scenario segmentation allows the model to better capture the characteristics of data in each scenario, thereby improving prediction accuracy. Without scenario segmentation, the model's training data contains more heterogeneity, which interferes with the model's learning process. Power generation is affected by weather conditions, resulting in varying data distributions. Segmenting weather scenarios helps the CRMN model better adapt to the inherent structure of the data, thus improving its performance across various evaluation metrics.

[0090] This embodiment also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0091] This embodiment also provides a storage medium on which a computer program is stored, which, when executed by a processor, implements the method described thereon.

[0092] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather prediction information and scene perception strategies, characterized in that, Includes the following steps: Historical power data and numerical weather forecast data of photovoltaic power plants are preprocessed to obtain preprocessed photovoltaic power data and meteorological variable data. Based on the preprocessed photovoltaic power data and meteorological variable data, the correlation between each meteorological variable and photovoltaic power was analyzed using the Pearson correlation coefficient to screen out key meteorological factors. The key meteorological factors are input into a Gaussian mixture model for unsupervised clustering to construct a sample set of typical weather scenarios. Based on the typical weather scenario sample set, for each typical weather scenario, a corresponding convolutional residual memory network prediction model is trained to obtain several sub-models. The aforementioned sub-models are fused together to output the day-ahead power generation prediction results for photovoltaic power plants.

2. The method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy according to claim 1, characterized in that, The key meteorological factors include: air temperature, total radiation, and cloud cover.

3. The method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy according to claim 1, characterized in that, The probability density function of the Gaussian mixture model is expressed as: ； In the formula, It is the true distribution of the whole. The estimate, Let be the weighting coefficients, mean, and variance of the k-th Gaussian component, respectively. Let be the k-th Gaussian component of the GMM.

4. The method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy according to claim 3, characterized in that, The convolutional residual memory network prediction model includes: a convolutional neural network, a residual connection mechanism, and a gated recurrent unit; The process of outputting the power prediction result based on the convolutional residual memory network prediction model includes: A convolutional neural network is used to extract spatial features from the input meteorological feature sequence to obtain high-dimensional spatial features; The high-dimensional spatial features are fused using a residual connection mechanism to obtain an enhanced feature sequence. A gated recurrent unit is used to model the temporal dependency relationship of the enhanced feature sequence to obtain a temporal feature representation; The photovoltaic power generation prediction results are obtained by performing regression processing on the time-series feature representation based on the fully connected layer.

5. The method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy according to claim 4, characterized in that, The hidden layer output expression of the gated loop unit is: ； ； ； ； In the formula, r t To reset the door, z t To update the gate, h t Currently in a hidden state. It is the Sigmoid activation function. It is the hyperbolic tangent function. For the current input, The state was hidden in the previous moment. In the candidate hidden state, and These represent resetting the gate weights and updating the gate weights, respectively. * represents the Hadamard product. The weight coefficients represent the candidate hidden states.

6. The method for predicting the day-ahead power generation of a photovoltaic power plant by integrating numerical weather forecast information and scene perception strategy according to claim 1, characterized in that, The process of obtaining several sub-models includes: testing the corresponding training convolutional residual memory network prediction model through several evaluation metrics to obtain the optimal sub-model; The evaluation indicators include: root mean square error, mean absolute error, mean error, coefficient of determination, pass rate, and accuracy.

7. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.

8. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 1.

Images