A method for predicting power generation of a photovoltaic power generation system

Abstract

The application provides a power generation amount prediction method of a photovoltaic power generation system, and relates to the technical field of photovoltaic power generation, and comprises the following steps: performing data collection of the photovoltaic power generation system, and establishing a feature data set; a functional model agent pool is established, the prediction model in the pool corresponds to the feature data set one by one, and the prediction model is a weak prediction model; a prediction benefit function is configured, and the evaluation dimensions include prediction accuracy, confidence distribution, and correlation influence contribution; the corresponding prediction model is called to perform power generation amount prediction, and the prediction game is performed by using the benefit function; continuous iteration is performed, the output weight is updated, and the cooperation strategy is adjusted; when iteration stops at any time, the corresponding output weight and cooperation strategy are read, and a power generation amount prediction result is constructed. The application solves the technical problem that the prior art usually depends on a single prediction model to perform photovoltaic power generation amount prediction, cannot fully capture various influence factors in the system, and leads to poor prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power generation, in particular to a power generation capacity prediction method of a photovoltaic power generation system. BACKGROUND

[0002] As a green and environmentally friendly energy form, photovoltaic power generation is widely used. Photovoltaic power generation system converts solar energy into electrical energy, providing a sustainable solution for energy supply. However, the power generation capacity of photovoltaic power generation system is affected by many factors, such as weather, equipment status, seasonal changes, geographical location, etc. Therefore, accurately predicting the power generation capacity of photovoltaic power generation system is crucial for power grid dispatching, energy management and system maintenance.

[0003] However, the existing technology usually relies on a single prediction model for photovoltaic power generation capacity prediction. The single model cannot fully capture various influencing factors in the system. Although ensemble learning methods try to improve prediction accuracy through multi-model fusion, in practical applications, existing multi-model methods do not fully utilize the complementary advantages between different models, often lacking effective inter-model collaboration strategies. Therefore, these methods ignore the complementarity between models, leading to unstable prediction results, especially in complex or extreme environments, the prediction accuracy is poor. SUMMARY

[0004] The present application provides a power generation capacity prediction method of a photovoltaic power generation system, aiming to solve the technical problem that the existing technology usually relies on a single prediction model for photovoltaic power generation capacity prediction, which cannot fully capture various influencing factors in the system, and the existing multi-model method cannot fully utilize the complementary advantages between different models, lacking effective inter-model collaboration strategies, resulting in poor prediction accuracy.

[0005] The power generation capacity prediction method of a photovoltaic power generation system disclosed in the present application, the method comprises: performing data collection of a photovoltaic power generation system, establishing a feature data set, the feature data set includes environmental feature data set, equipment feature data set, time feature data set, spatial feature data set; Establish a functional model agent pool, the prediction model in the functional model agent pool corresponds to the feature data set one by one, and the prediction model is a weak prediction model; Configure a prediction benefit function, the evaluation dimensions of the prediction benefit function include prediction accuracy, confidence allocation, and associated influence contribution; After inputting the feature data set into the functional model agent pool, the corresponding prediction model is called to perform power generation capacity prediction, and the benefit function is used for prediction game; Perform continuous iteration of prediction game, update the output weight of each prediction model, and adjust the collaboration strategy; When the iteration at any time stops, read the corresponding output weight and collaboration strategy, and construct the power generation capacity prediction result.

[0006] The one or more technical solutions provided in the application have at least the following beneficial effects:

[0007] By comprehensively collecting various data of the photovoltaic power generation system, a multi-dimensional feature data set is established, which provides rich and diverse input data for the power generation prediction, can comprehensively reflect various factors affecting the power generation, and improves the prediction accuracy; by creating a functional model agent pool, each prediction model focuses on processing specific data features, so that each model can be optimized for different data features, which improves the flexibility and adaptability of the prediction system; the use of weak prediction models in the model pool enables the system to take advantage of multiple lightweight models, avoids the high computational cost of a single complex model, and reduces the resource occupation of the system; the configuration of the benefit function ensures multi-dimensional evaluation of the model performance by introducing multiple evaluation dimensions, which enables each model to adjust not only according to its accuracy but also according to its contribution to the final prediction and cooperation relationship, thereby improving the overall prediction effect; through prediction game, multiple models can optimize their outputs through cooperation and competition mechanism, and in the game process, the performance of the models is evaluated by the benefit function, which not only dynamically adjusts the weight of each model but also maximizes the synergistic effect between models, thereby improving the overall performance of the prediction; through continuous iteration and optimization, the weight of each model is automatically adjusted according to its performance after each round of game, which enables the model weight to gradually converge to the optimal value, reduces the prediction error of the system, and through adjusting the cooperation strategy between models, the cooperation between multiple models can be made more closely, avoiding the over-reliance on a single model, and ensuring that each model plays its due role in the power generation prediction; by reading the output weight and cooperation strategy when the iteration stops, a stable and optimal prediction result can be obtained, which means that the best model combination and cooperation strategy have been found in multiple iterations, which can provide more accurate and reliable power generation prediction, and this process effectively avoids unnecessary calculation and complexity in the prediction process, enabling the system to efficiently obtain the final power generation prediction result while ensuring high prediction accuracy.

[0008] The above description is only a summary of the technical solutions of the application. In order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 A power generation prediction method flowchart of a photovoltaic power generation system is provided for the embodiments of the application.

[0010] Figure 2A power generation prediction method of a photovoltaic power generation system is provided in the embodiments of the present application. DETAILED DESCRIPTION

[0011] The embodiments of the present application provide a power generation prediction method of a photovoltaic power generation system, which solves the technical problem in the prior art that a single prediction model is usually used for photovoltaic power generation prediction, various influencing factors in the system cannot be fully captured, the complementary advantages between different models cannot be fully utilized in the existing multi-model method, there is a lack of effective inter-model collaboration strategy, and the prediction accuracy is poor.

[0012] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be specifically introduced in combination with the drawings of the specification.

[0013] As shown in the drawings, Figure 1 The embodiments of the present application provide a power generation prediction method of a photovoltaic power generation system, which comprises:

[0014] Data collection of the photovoltaic power generation system is performed, a feature data set is established, and the feature data set comprises an environmental feature data set, a device feature data set, a time feature data set, and a space feature data set.

[0015] Various data of the photovoltaic power generation system are collected to establish a comprehensive feature data set, and the feature data set comprises multiple subsets. Specifically, the environmental feature data set refers to various factors of the environment where the photovoltaic system is located, such as weather conditions, temperature, humidity, solar radiation intensity, wind speed, air quality, etc. The data can be collected in real time through sensors or obtained from a meteorological data source. The device feature data set refers to various hardware state information inside the photovoltaic power generation system, such as the output power of the photovoltaic panel, the working state of the inverter, the power loss, the health status of the components, etc. These data are collected through sensors, monitoring systems, and other devices installed in the system, and involve performance monitoring data of the equipment. The time feature data set refers to factors related to time, such as the time period of system operation, season, distinction between weekdays and weekends, and length of illumination time. The time feature is particularly important for the prediction of the photovoltaic power generation system, because the illumination intensity changes with time, especially seasonal changes. The space feature data set involves the features of the installation location of the photovoltaic system, such as geographic coordinates (latitude and longitude), orientation, and inclination angle, etc. These space features directly affect the amount of solar energy received, because the angle and position of the photovoltaic system determine the efficiency of solar radiation reception. By collecting these feature data and organizing them into a structured feature data set, rich input data can be provided for subsequent prediction models.

[0016] A functional model agent pool is established, and the prediction models in the functional model agent pool correspond one-to-one to the feature data set, and the prediction models are weak prediction models.

[0017] A functional model agent pool is established, which refers to a centralized model management pool containing multiple prediction models based on different algorithms or frameworks such as linear regression, decision tree, support vector machine, neural network, etc., or other models suitable for time series prediction, regression analysis, etc. These prediction models are weak prediction models, meaning they are not particularly strong individually, but through integration methods, they can improve the accuracy of prediction. Each prediction model corresponds to a different part of the feature data set, for example, some prediction models focus on environmental feature data sets, while others focus on device features or time features. In the agent pool, a one-to-one correspondence is established between the prediction model and the feature data set, which helps to optimize different types of data.

[0018] A prediction benefit function is configured, and the evaluation dimensions of the prediction benefit function include prediction accuracy, confidence allocation, and correlation impact contribution.

[0019] A prediction benefit function is configured to evaluate the output of each prediction model to optimize the prediction model in subsequent prediction game processes. This prediction benefit function evaluates the performance of each model through multiple dimensions, including prediction accuracy, confidence allocation, and correlation impact contribution. Prediction accuracy refers to the gap between the model's predicted results and the actual results. Typically, prediction accuracy is measured by calculating errors such as mean squared error or absolute error. A smaller error value means higher accuracy. Confidence allocation is the degree of confidence that a model has in its prediction results, indicating the degree to which the model believes its prediction is reliable under a certain prediction result. Higher confidence means the model is more confident in its prediction. Confidence can be represented by probability distribution or confidence interval, such as regression models that can output a predicted value and its confidence interval, indicating the model's confidence in the prediction result. High-confidence prediction results will be given higher weights in decision-making. Correlation impact contribution measures the contribution of each prediction model to the power generation prediction result, especially the importance of each feature in prediction. Feature importance evaluation techniques can be used to evaluate the impact of each feature on the prediction result. In subsequent prediction game processes, features with greater impact are given higher weights.

[0020] After inputting the feature data set into the functional model agent pool, the corresponding prediction model is called to predict power generation, and the benefit function is used for prediction game.

[0021] The complete feature dataset is input into the functional model agent pool, and each prediction model is trained for a specific feature dataset in the agent pool, so according to the input feature dataset, the model prediction calculation is performed to obtain the corresponding power generation prediction result. The prediction game is an optimization process, in which each prediction model evaluates its performance in prediction according to the comparison between its output result and the actual result through the profit function. In the prediction game, the three evaluation dimensions of the profit function (prediction accuracy, confidence allocation, and correlation impact contribution) will determine how to adjust the output weight of each model. For example, if the prediction accuracy of a certain model is high and its confidence is large, its output result will be given a higher weight, and vice versa, if the prediction error of the model is large, it will be given a smaller weight in the next round of game.

[0022] The continuous iteration of the prediction game is performed, the output weight of each prediction model is updated, and the collaboration strategy is adjusted.

[0023] The continuous iteration process of the prediction game is a repeated optimization process. In each iteration, the performance of each model is evaluated using the prediction profit function according to the current output of each prediction model. According to the accuracy, confidence, and correlation impact contribution of the prediction result, the output weight of each model is updated, that is, the influence degree of each model on the final power generation prediction result is adjusted. Specifically, if a certain model provides more accurate prediction in the current iteration (e.g., lower prediction error or higher confidence), its output weight is increased; if the prediction performance of a certain model is poor (e.g., high error or low confidence), its output weight is reduced. This process ensures that more accurate models will play a greater role in the next round of prediction, while models with poor performance will gradually reduce their influence.

[0024] In each iteration, not only the weight of a single model is updated, but also the collaboration strategy among models is adjusted, which means that according to the mutual influence, collaboration degree and relative importance among different models, the roles of them in power generation prediction are dynamically adjusted. The adjustment of the collaboration strategy means the optimization of the degree of information sharing and collaboration among different prediction models. For example, if the collaboration between two models is good (i.e., their prediction results are complementary), the collaboration between them will be increased; on the contrary, if the collaboration between some models is poor, the collaboration strategy between them will be weakened or reduced. The adjustment of the collaboration strategy helps to reduce redundant prediction and optimize the collaboration relationship among models, so that the system can more accurately predict power generation after multiple iterations.

[0025] When the iteration at any time stops, the corresponding output weight and collaboration strategy are read to construct the power generation prediction result.

[0026] The iteration process ends when a certain stop condition is met, common stop conditions include maximum number of iterations, error convergence, profit function convergence, etc. When any stop condition is met, the iteration process will terminate. After the iteration stops, the final output weight of each prediction model and the collaboration strategy are read, where the final weight of each model reflects its performance in the iteration process, and the model with better performance will have a higher weight, while the model with lower weight will be lower. In the multi-model collaboration framework, the final collaboration strategy determines the interaction between models, such as the way of information sharing, how to weight, etc. According to the output weight and the collaboration strategy, the power generation prediction result is constructed. Specifically, according to the weight of each model, the weighted average value is calculated as the final prediction value. If there are multiple prediction model results, combine their prediction outputs to generate the final prediction using the weight adjustment fusion strategy. Combine the collaboration strategy to further adjust the prediction results of different models to optimize the overall performance. The final power generation prediction result can be used to make actual power generation management decisions, such as resource scheduling, load prediction, equipment maintenance, etc.

[0027] Further, as shown in Figure 2 The prediction game using the profit function includes:

[0028] Establish the initial weight of each prediction model. After inputting the feature data set into the prediction model, use the prediction model to predict the power generation, and evaluate the profit value of each power generation prediction result through the profit function. Adjust the output weight of each prediction model based on the initial weight using the profit value, and adjust the collaboration strategy between models to complete a round of update iteration.

[0029] Establish the initial weight of each prediction model, which reflects the influence of the model on the prediction result in the initial stage. The selection of the initial weight can be based on multiple factors, such as the historical performance of the model, the type of the model, the coverage of the feature data, etc. For example, if a certain model performs well in the past prediction task, it can be assigned a higher initial weight.

[0030] Input the feature data set into the prediction model, each model processes the corresponding feature data set and performs power generation prediction calculation. The prediction result is usually a continuous numerical value representing the power generation under given conditions. The prediction result of each model is evaluated by a predefined profit function, including prediction accuracy, confidence allocation, and associated impact contribution. Each model's prediction result is assigned a profit value after being evaluated by the profit function, indicating the model's effectiveness in the current prediction.

[0031] The initial weights of the prediction models are adjusted based on their payoffs, with better-performing models receiving higher weights and poorer-performing models receiving lower weights. Adjusting the collaborative strategy involves optimizing information sharing and collaboration mechanisms between different models so that they complement each other rather than duplicate work. For example, if the prediction results of two models are highly complementary (i.e., their errors differ), the collaborative relationship between them should be strengthened to increase information sharing. In each round of the game, the payoffs of each model are continuously evaluated, and the model weights and collaborative strategies are adjusted based on these payoffs. This process helps to continuously optimize prediction accuracy and maximize the overall system's prediction effect by adjusting the relative importance of the models and the collaborative mechanisms. The end of an iteration means that the weights and collaborative strategies of all models have been updated, and preparations are made for the next round of predictions.

[0032] Furthermore, adjusting the output weight of each prediction model based on the initial weight using the benefit value includes:

[0033] The output weight is adjusted through the weight update formula as follows:

[0034] ;

[0035] in, Characterization At the first iteration, The output weight of the prediction model, Characterization At the first iteration, The output weight of the prediction model, The main driving factor of returns is For the The prediction model in The profit value of the iteration, Characterization in the The average return of all prediction models in iterations, is the collaborative feedback factor, Representation is different from predictive model The prediction model, Characterization in the The prediction model under the iteration and prediction models The cooperative mutual benefit coefficient is used to measure the predicted synergistic relationship between them. Characterization At the first iteration, The output weights of the prediction model.

[0036] Specifically, the weight update formula is as follows:

[0037] ;

[0038] wherein, characterizing the degree of cooperation and mutual benefit coefficient of the prediction model at the first iteration and the prediction model , used to measure the prediction synergy relationship between them, calculated by the following factors: output correlation (such as Pearson correlation coefficient), joint lift rate (relative to the degree of improvement of individual prediction when used together), Bayesian contribution or Shapley approximation value.

[0039] The formula is used for weight update of the prediction model, and the core idea is to continuously optimize the weight of each model, so that the overall prediction result is more accurate. In each iteration, the weight of the model is determined by three factors, wherein, is the initial weight term of each iteration; is the benefit-driven part of the model, which adjusts the weight according to the prediction benefit of each model and the average benefit difference of all models, ensuring that the prediction result of the model is closer to the overall trend; is the cooperation part, which adjusts the complementary action between each model based on the cooperation effect between models to promote cooperation between models and reduce the difference in weight between models.

[0040] Overall, the formula optimizes the weight of the prediction model through the two mechanisms of benefit-driven and cooperation adjustment. In each iteration, the weight of the model is not only adjusted according to its own performance (benefit), but also affected by the output of other models. Through the adjustment of weight difference and cooperation effect, the overall prediction effect is finally optimized.

[0041] Further, the output weight adjustment by the weight update formula includes:

[0042] A smoothing penalty mechanism is established to perform historical fluctuation analysis of the prediction model. Trigger analysis of the smoothing penalty mechanism is performed according to the historical fluctuation analysis result. A first weight evolution stable feedback is generated using the trigger analysis result. A memory update mechanism is established. If the historical prediction result of the prediction model is a stable result and the current iteration prediction is abnormal, the memory update mechanism is triggered. A second weight evolution stable feedback is generated using the triggered memory update mechanism. The output weight adjustment is constrained according to the first weight evolution stable feedback and the second weight evolution stable feedback.

[0043] A smoothing penalty mechanism is established to control the change of the weight of the prediction model in multiple iterations. If the weight change of a certain model is too large, the smoothing penalty mechanism will limit such fluctuations, thereby avoiding unstable phenomena.

[0044] Through historical volatility analysis, the differences between the historical prediction results and the actual results of each model are calculated, and the volatility of these differences reflects the stability and accuracy of the model. The volatility of the model output can be evaluated by calculating the error standard deviation or the root mean square error of each iteration. Higher volatility means that the model's prediction is unstable, and lower volatility indicates that the model performs more stably. When the historical volatility reaches a preset threshold, the smoothing penalty mechanism is triggered. Specifically, models with high historical volatility are penalized, and their weight fluctuations in prediction are reduced. The triggering analysis of this mechanism can be performed by setting a volatility tolerance, for example, if the error standard deviation of the model exceeds a certain range, the penalty is executed.

[0045] Based on the historical volatility analysis and the triggered smoothing penalty mechanism, the first weight evolution stability feedback is generated, which is used to guide the model to adjust the weights in the next iteration, so that the weight change of the model is more stable, and the uncertainty caused by excessive volatility is reduced. Generally, this feedback is achieved by adjusting the update speed or intensity of the model weights, for example, by reducing the update rate to avoid large weight adjustments.

[0046] A memory update mechanism is established to store and update the historical prediction model state, especially when the model's prediction results in a certain iteration are abnormal, this mechanism can help the model recover to a more stable state. Prediction abnormalities can be defined by large prediction errors, abnormal model outputs, or significant deviations from actual data. For example, when the model's historical prediction results are relatively stable, and the current prediction results differ greatly from the historical results, the memory update mechanism is triggered. This mechanism will refer to the past performance of the model to correct the current prediction error.

[0047] The second weight evolution stability feedback is generated according to the memory update mechanism, which is used to correct the current weight adjustment to make it more consistent with the stability of the historical performance. The form of this feedback includes adjusting the current prediction results back to a similar level as the historical prediction, or resetting the weights by weighted averaging of the historical stable results.

[0048] The first feedback comes from the analysis of the smoothing penalty mechanism, focusing on the stability of the prediction model in historical volatility, and smoothing the model's output by limiting weight fluctuations. The second feedback comes from the memory update mechanism, focusing on the recovery process of the model when it makes abnormal predictions, and restoring its historical stable state by adjusting the model weights. These two feedback mechanisms work together to adjust the model weights to ensure that the model's output is more stable and not affected by abnormal predictions or excessive fluctuations. Finally, after the influence of the smoothing penalty and memory update mechanisms, the final output weights of each prediction model are generated, which will be used in subsequent iterations to ensure that the model is more stable and accurate in the next round of prediction.

[0049] Further, the adjustment of the collaboration strategy includes:

[0050] Establishing a benefit feedback feature including individual benefit feedback and average benefit deviation; establishing a prediction result correlation feature including an output similarity index and a prediction coupling degree index; and establishing a scenario label dominant feature, and adjusting the collaboration strategy based on the benefit feedback feature, the prediction result correlation feature, and the scenario label dominant feature.

[0051] A benefit feedback feature is constructed to evaluate the performance of each prediction model and optimize it. The benefit feedback feature mainly includes individual benefit feedback and average benefit deviation. The individual benefit feedback refers to the benefit of the prediction result of each prediction model in the current iteration, which is calculated by a benefit function according to the prediction accuracy, confidence allocation, and contribution degree of the model to evaluate the performance of the model in the iteration. The individual benefit feedback reflects the contribution degree of a single model and reflects the prediction quality of the model. If the prediction of a model is very accurate, its benefit feedback will be higher. If the prediction error of the model is larger, its benefit feedback will be lower. The average benefit deviation measures the deviation between the average benefit of the current all models and the average benefit of the last iteration. This index reflects the overall performance change of all models. If the benefit of the current iteration is significantly higher than that of the last iteration, it means that the overall performance of the models has improved, and vice versa. The average benefit deviation helps to evaluate the improvement of the system in multiple iterations. If the collective performance of the models is not as expected, the collaboration of the models can be optimized by adjusting the strategy.

[0052] A prediction result correlation feature is established to measure the relationship between different prediction models to further optimize the collaboration strategy. The prediction result correlation feature includes an output similarity index and a prediction coupling degree index. The output similarity index measures the similarity between the outputs of different prediction models. This index is achieved by calculating the similarity between the prediction results of different models, such as using cosine similarity, Pearson correlation coefficient, etc. Models with high output similarity represent similar ways of processing specific data features, so their prediction results may have high redundancy. In this case, the collaboration strategy is adjusted to reduce the redundancy of these models and improve system efficiency. The prediction coupling degree index measures the collaboration and complementarity between multiple models. It measures the close relationship between the prediction results of different models by quantifying the correlation or synergy between model outputs. If the coupling degree between two models is high, it means that they provide valuable complementary information for the same prediction task. In this case, the collaboration between the two models is enhanced, and their influence in the final prediction is increased.

[0053] Establishing scene tag dominant features, which refers to assigning a scene tag to each prediction task according to the specific scene or environment of different prediction tasks. The scene tag reflects the specific requirements of the current task, such as weather conditions, time periods, device status, and other factors. These scene tags help the system adjust the collaboration strategy according to the needs of different tasks to ensure that the model performs optimally in specific scenarios. For example, in extreme weather conditions, some models are better at handling environmental features, while other models are more accurate in device health status.

[0054] Adjusting the collaboration strategy based on the revenue feedback feature, the prediction result correlation feature, and the scene tag dominant feature. For example, when some models perform well in specific scenarios, the weights of these models can be increased to enhance their influence in the final prediction. If the output similarity between models is high and the prediction coupling degree is strong, the influence of these models can be reduced to prevent redundancy. Conversely, if some models perform poorly in specific scenarios, the collaboration strategy can be adjusted to reduce their participation or even temporarily disable them. Based on the comprehensive analysis of the above features, the collaboration mode of each model is dynamically adjusted, including changing the weight distribution between models, adjusting the information sharing mechanism, and optimizing the complementarity between models. This strategy adjustment helps the system maximize prediction performance in different scenarios and task requirements.

[0055] Furthermore, the prediction game using the revenue function further includes:

[0056] Creating an extreme scene recognition channel, inputting the feature data set into the extreme scene recognition channel, performing abnormal recognition of extreme weather, abnormal device status, and spatial mutation events, and establishing a scene recognition tag. Using the scene recognition tag to perform micro-model hot loading, and adding the hot-loaded micro-model to the prediction game to perform scene enhancement management.

[0057] Creating an extreme scene recognition channel, which is used to analyze input data and detect abnormal situations. It receives various information from the feature data set, such as environmental features, device status, and spatial changes, and performs specific recognition tasks on these data. The channel uses machine learning models, rule engines, or deep learning networks to analyze the data and detect possible extreme events in real time. For example, based on weather data, it identifies extreme weather such as strong winds, heavy rain, and snowstorms; based on device data, it identifies device failures or abnormal operating conditions; and based on spatial data, it identifies sudden events caused by changes in geographical or spatial location, such as shading and device damage. According to the above analysis, corresponding scene recognition tags are generated, which represent the types of extreme events in the current data, such as heavy rain, device failure, and shading effect. Scene tags identify the current abnormal state and provide a basis for subsequent processing steps.

[0058] Micro-model refers to small, lightweight models optimized for specific scenarios, which do not require global training, but are quickly deployed for specific abnormal scenarios, for example, in extreme weather conditions, load a micro-model specifically for weather impact, hot loading means dynamically loading new micro-models without restarting the entire system, this mechanism ensures that the system can quickly respond to new scenario changes in real-time environment. When an abnormal scenario is identified through the extreme scenario identification channel and a scenario identification label is generated, the corresponding micro-model is quickly loaded according to the label, for example, if a device failure scenario label is identified, a micro-model specifically for device failure is loaded to adjust the prediction strategy.

[0059] When the micro-model is hot-loaded and ready, it will participate in the prediction game, which means that these micro-models are not used alone, but are used in collaboration and game with other models to improve the prediction ability of the entire system. Specifically, the output of the micro-model is integrated into the prediction system, together with other existing prediction models, and optimized based on the revenue function to ensure accurate prediction in abnormal scenarios. Scenario enhancement management refers to adjusting the collaboration strategy between models after identifying specific abnormal scenarios, so that micro-models and other main models can work together to maximize prediction performance. In this way, the collaboration method of the model can be flexibly adjusted when facing different scenarios and changes to ensure the accuracy of the prediction result.

[0060] Further, the adding of the hot-loaded micro-model to the prediction game and the performing of scenario enhancement management include:

[0061] Configuring an output intervention threshold; when performing scenario enhancement management, using the output intervention threshold to identify intervention trigger of the micro-model; using the intervention trigger identification result to manage the prediction game.

[0062] The output intervention threshold refers to triggering the intervention mechanism when the output result of the model exceeds or is lower than a certain preset range during the prediction process. These thresholds can be defined based on prediction error, confidence of prediction result or deviation of model output. Too low threshold may lead to excessive intervention, while too high threshold may lead to system delay, so accurate adjustment is needed.

[0063] When performing scenario enhancement management, first identify the current abnormal situation according to the scenario label, such as extreme weather, device failure, etc., then compare the prediction result of the current micro-model with the output intervention threshold, if the output of the micro-model exceeds the preset intervention threshold, the intervention mechanism will be triggered to start adjusting or ignoring the prediction result of the model.

[0064] According to the intervention trigger identification result, the model weights in the game are adjusted, if some models are marked as abnormal, their weights are reduced or completely removed, by adjusting the game strategy, the weights are redistributed, the weights of the models with reliable prediction results and stable performance are increased, so as to improve the prediction accuracy.

[0065] Further, the performing the continuous iteration of the prediction game further comprises:

[0066] A low yield threshold is established, after calculating the prediction model yield value of each iteration, the trigger identification of the low yield threshold is performed, if the low yield threshold trigger of any prediction model in the preset iteration window satisfies the expected proportion, an initialization instruction is generated, and the parameter initialization management of the corresponding prediction model is performed according to the initialization instruction.

[0067] The low yield threshold refers to that when the yield value of a certain prediction model is lower than a certain set value, it is considered that the prediction performance of the model is poor, and intervention is needed, the low yield threshold is set through historical data analysis or priori knowledge. In each iteration, the prediction yield of each model is calculated according to the yield function, the yield value includes the prediction accuracy, confidence allocation, contribution degree and other indicators of the model, the yield value of each model is tracked, and it is judged whether it is lower than the set low yield threshold, if the yield value of a certain model is lower than the preset low yield threshold, the low yield threshold trigger identification is triggered, which means that the model fails to achieve the expected performance level in the current iteration period, and needs to be further adjusted.

[0068] The preset iteration window is a fixed time range or iteration number, which is used to evaluate the performance of the model in continuous several iterations, this window helps to understand the long-term performance of the model, rather than the fluctuation of single iteration, if the low yield threshold of any model is triggered more than a certain proportion in this iteration window, it is considered that the model fails to adapt to the current data or task for a long time, and needs to be reinitialized.

[0069] The expected proportion refers to that the proportion of the low yield threshold trigger exceeds a certain predetermined threshold in the preset iteration window, for example, if the yield value of the model of 2 rounds is lower than the low yield threshold in the iteration window of 5 rounds, the trigger proportion is 40%, if the proportion exceeds the set expected proportion (such as 50%), the initialization instruction is generated, the expected proportion can be flexibly adjusted according to the fault tolerance ability of the model, the task demand and the business scene. When the low yield threshold satisfies the expected proportion, the initialization instruction is automatically generated, which is used to start the reinitialization process of the model, so as to ensure that the model can recover to a better prediction state.

[0070] Parameter initialization refers to resetting the parameters of the model to ensure that the model can adapt to the data and task requirements when the model performs poorly. Parameter initialization includes weight reset, etc. After initialization management, the reinitialized model is monitored to ensure that its performance improves, and the parameters are adjusted in a timely manner.

[0071] Further, if the low yield threshold of any prediction model within the preset iteration window triggers the desired proportion, an initialization instruction is generated, including:

[0072] Activate the sliding window to perform residual accumulation analysis on the prediction model and generate a drift trend prediction result. Establish an auxiliary authentication based on the drift trend prediction result and trigger compensation through the auxiliary authentication for the desired proportion.

[0073] The sliding window is a fixed-length time window used for local analysis of the model's output. Each time the window is moved, new data points are considered, and the oldest data points are discarded. The sliding window helps dynamically evaluate the model's performance by analyzing the model's error changes, i.e., residuals, in the short term, and timely detects the model's prediction deviation and trends. Residuals refer to the difference between the model's predicted value and the actual observed value, i.e., the difference between the model's predicted value and the actual observed value.

[0074] In each sliding window, the residual of each prediction model is calculated, and these residuals are analyzed cumulatively. Through the cumulative change of residuals, the stability and accuracy of the model can be evaluated. If the model's residuals continue to increase or become unstable in multiple windows, the prediction trend of the model is identified to be deviating, thereby providing a basis for subsequent adjustments. The drift trend prediction result is obtained by analyzing the residual change trend in the sliding window to predict the future drift direction. For example, if the residuals in consecutive windows gradually increase, it indicates that the model's performance will continue to decline, and vice versa, indicating that the model is becoming more stable.

[0075] An auxiliary authentication is established using the generated drift trend prediction result. In the auxiliary authentication, the relationship between the drift trend prediction result and the expected prediction result is compared to determine whether the model needs to be further adjusted. If the drift trend prediction result indicates that a certain model consistently deviates, the auxiliary authentication mechanism confirms whether the model needs compensation. According to the auxiliary authentication result, if the drift trend prediction result indicates that the model's prediction result significantly deviates from the actual demand and this deviation meets the set desired proportion, compensation measures are triggered. According to the compensation strategy, the model's weights, parameters, or other factors are adjusted to correct the deviation. The purpose of compensation is to ensure that the system's prediction performance returns to the ideal state and reduces the error of deviating from the true result.

[0076] Further, after constructing the power generation prediction result, the following steps are included:

[0077] The power generation task demand is obtained, and matching analysis is performed according to the power generation task demand and the power generation prediction result to establish supply and demand abnormalities. The supply and demand abnormalities are used to report abnormal early warnings, and abnormal management is performed through the abnormal early warnings.

[0078] The power generation task demand is the power generation required according to a specific time period, region or task requirement, which is usually determined by a user, a dispatching system or external factors. The power generation prediction result is the future power generation obtained through a prediction model. The power generation task demand and the power generation prediction result are matched and analyzed, specifically, the difference between the predicted power generation and the actual demand is compared to analyze whether the demand can be met. If the predicted power generation is lower than the demand and the difference exceeds a preset threshold, a supply shortage is identified. If the predicted power generation is higher than the demand and the difference exceeds a preset threshold, an excess is identified. Based on the difference between the prediction result and the task demand, supply and demand abnormalities are established, including supply shortage and resource excess.

[0079] When the supply and demand abnormalities are identified, an abnormal early warning mechanism is triggered, and the content of the abnormal early warning includes the abnormal type, the abnormal degree, the influence range and the like. Abnormal management is performed according to the abnormal early warning. The abnormal management refers to a series of measures taken to solve the imbalance between supply and demand after receiving the abnormal early warning. The management measures include adjusting the power generation plan, adjusting the load distribution, optimizing the resource dispatching, coordinating external resources and the like. Through the abnormal management, it is ensured that the system can timely identify and respond to the imbalance between power supply and demand, so as to optimize the configuration and dispatching of power resources and ensure the stable operation of the power system.

[0080] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method of predicting power generation of a photovoltaic power generation system, characterized by, The method comprises: performing data collection of a photovoltaic power generation system, establishing a feature data set, the feature data set comprising an environmental feature data set, a device feature data set, a time feature data set, and a spatial feature data set; establishing a functional model agent pool, the prediction models in the functional model agent pool corresponding one-to-one to the feature data set, and the prediction models being weak prediction models; configuring a prediction revenue function, the evaluation dimensions of the prediction revenue function including prediction accuracy, confidence allocation, and associated impact contribution; after inputting the feature data set into the functional model agent pool, calling the corresponding prediction models to perform power generation capacity prediction, and using the revenue function to perform prediction game; performing continuous iteration of the prediction game, updating the output weight of each prediction model, and adjusting the cooperation strategy; when the iteration at any time stops, reading the corresponding output weight and cooperation strategy to construct a power generation capacity prediction result; the adjustment of the cooperation strategy comprises: establishing a revenue feedback feature, the revenue feedback feature comprising individual revenue feedback and average revenue deviation, the individual revenue feedback referring to the revenue of the prediction result of each prediction model in the current iteration, and the average revenue deviation measuring the deviation between the average revenue of all models in the current iteration and the average revenue of the last iteration; establishing a prediction result association feature, the prediction result association feature comprising an output similarity index and a prediction coupling degree index, the output similarity index measuring the similarity between the outputs of different prediction models, and the prediction coupling degree index measuring the cooperation and complementarity between multiple models; establishing a scenario label dominant feature, adjusting the cooperation strategy based on the revenue feedback feature, the prediction result association feature, and the scenario label dominant feature, the scenario label dominant feature referring to assigning a scenario label to each prediction task according to the specific scenario or environment of different prediction tasks; the use of the revenue function to perform prediction game comprises: establishing an initial weight of each prediction model; after inputting the feature data set into the prediction model, using the prediction model to perform power generation capacity prediction, and evaluating the revenue value of each power generation capacity prediction result through the revenue function; using the revenue value to adjust the output weight of each prediction model based on the initial weight, and adjusting the cooperation strategy between models to complete an update iteration.

2. The method of claim 1, wherein the power generation of the photovoltaic power generation system is predicted based on the solar radiation amount, the temperature, and the wind speed. the use of the revenue value to adjust the output weight of each prediction model based on the initial weight comprises: adjusting the output weight through a weight update formula, as follows: ； wherein, characterizing the output weight of the first prediction model at the first iteration, characterizing the output weight of the first prediction model at the first iteration, is a benefit main driver factor, is a benefit value of the first prediction model at the first iteration, characterizing the benefit average value of all prediction models at the first iteration, is a collaborative feedback factor, characterizing a prediction model different from the prediction model , characterizing the collaborative mutual benefit degree coefficient of the prediction model and the prediction model at the first iteration, for measuring the prediction synergy relationship therebetween, characterizing the output weight of the first prediction model at the first iteration.

3. The method of claim 2, wherein the power generation of the photovoltaic power generation system is predicted by using the power generation prediction model. the adjustment of the output weight through the weight update formula comprises: establishing a smoothing penalty mechanism, performing historical fluctuation analysis of the prediction model, performing trigger analysis of the smoothing penalty mechanism according to the historical fluctuation analysis result, generating a first weight evolution stable feedback using the trigger analysis result; establishing a memory update mechanism, if the historical prediction result of the prediction model is a stable result and the prediction of the current iteration is abnormal, triggering the memory update mechanism, and generating a second weight evolution stable feedback using the triggered memory update mechanism; adjusting the output weight according to the first weight evolution stable feedback and the second weight evolution stable feedback.

4. The method of claim 1, wherein the power generation of the photovoltaic power generation system is predicted by using the power generation data of the photovoltaic power generation system and the weather data of the weather information system. The prediction game using the benefit function further includes: An extreme scenario identification channel is created, the feature data set is input into the extreme scenario identification channel, abnormal identification of extreme weather, abnormal equipment state, and spatial mutation event is performed, and a scenario identification label is established; The micro-model hot loading is performed using the scenario identification label; The hot-loaded micro-model is added to the prediction game, and scenario enhancement management is performed.

5. The method of claim 4, wherein the power generation of the photovoltaic power generation system is predicted by using the power generation prediction model. The hot-loaded micro-model is added to the prediction game, and scenario enhancement management is performed. An output intervention threshold is configured; When the scenario enhancement management is performed, the intervention trigger identification of the micro-model is performed using the output intervention threshold; The prediction game management is performed using the intervention trigger identification result.

6. The method for predicting power generation of a photovoltaic power generation system according to claim 1, wherein: The continuous iteration of the prediction game further includes: A low benefit threshold is established, and after the benefit value of the prediction model of each iteration is calculated, the trigger identification of the low benefit threshold is performed; If the low benefit threshold trigger of any prediction model in a preset iteration window satisfies an expected proportion, an initialization instruction is generated; According to the initialization instruction, the parameter initialization management of the corresponding prediction model is performed.

7. The method of claim 6, wherein the power generation of the photovoltaic power generation system is predicted by using the power generation prediction model. If the low benefit threshold trigger of any prediction model in a preset iteration window satisfies an expected proportion, an initialization instruction is generated, which includes: A sliding window is activated to perform residual accumulation analysis of the prediction model, and a deviation trend prediction result is generated; An auxiliary authentication is established using the deviation trend prediction result, and the trigger compensation of the expected proportion is performed through the auxiliary authentication.

8. The method for predicting power generation of a photovoltaic power generation system according to claim 1, wherein: After the power generation prediction result is constructed, the following is included: The power generation task demand is obtained, the matching analysis is performed according to the power generation task demand and the power generation prediction result, and the supply and demand anomaly is established; The abnormal early warning is reported using the supply and demand anomaly, and the abnormal management is performed through the abnormal early warning.

Images