Generated power prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of new energy, and provides a generation power prediction method and device, electronic equipment and a storage medium, and the method comprises the steps: inputting predicted meteorological data of a to-be-predicted power station in a target time period into a preferential model, and obtaining preferential probability distribution; the preferential probability distribution comprises a plurality of power prediction models and preferential conditions corresponding to the power prediction models; based on the preferential probability distribution, the target power prediction model is screened out from the plurality of power prediction models, and the target power prediction model is called to process the predicted meteorological data to obtain the generated power data, so that the model is not selected passively depending on historical back-test performance. Compared with the prior art, the method provided by the invention has the advantages that the optimal tool is selected according to the future predicted meteorological data, and the strategy of dynamically selecting the optimal tool according to the future weather well overcomes the defect of hysteresis in the traditional scheme, can better adapt to the sudden change of the weather mode, and improves the accuracy and stability of power generation power prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy, in particular to a power generation power prediction method and device, electronic equipment and storage medium. BACKGROUND

[0002] Although the proportion of wind energy and solar energy in the power system is rapidly increasing, they have intermittency, volatility and uncertainty, which brings great challenges to their use and popularization. Accurate power generation power prediction is a key technology to solve this problem, which can help the power grid to make scheduling plans in advance, optimize standby capacity, reduce curtailment of wind and light, and improve operational efficiency and economy.

[0003] At present, accurate power prediction is mostly achieved by a complex multi-stage process, that is, multiple results are first generated, and then backtesting or similar day optimization is used to determine the final result. Among them, backtesting is based on actual power generation power data of the station in the past period to evaluate the historical performance of each model and select the model with the smallest error for subsequent prediction, but it has a lag and cannot make forward-looking decisions. Similar day optimization is to find similar days in history with the help of the next day's weather forecast data, analyze the performance of the corresponding model to select the optimal one, and the result is suboptimal. Both methods are prone to unstable prediction results, affecting actual application. SUMMARY

[0004] The present application provides a power generation power prediction method, device, electronic equipment and storage medium to solve the problems of lag, suboptimality and instability existing in the existing backtesting and similar day optimization.

[0005] The present application provides a power generation power prediction method, comprising: obtaining prediction meteorological data of a power plant to be predicted in a target period; inputting the prediction meteorological data into an optimization model to obtain an optimization probability distribution output by the optimization model; the optimization probability distribution contains multiple power prediction models and their corresponding optimization conditions; Based on the optimization probability distribution, a target power prediction model is selected from the multiple power prediction models, and the target power prediction model is called to process the prediction meteorological data to obtain power generation power data of the power plant in the target period. Among them, the optimization model and the multiple power prediction models are trained based on sample meteorological data and sample power generation power data of sample power plants in a historical period.

[0006] According to the power generation power prediction method provided by the present application, the optimization condition includes the model weight value or the selection probability of the corresponding power prediction model. The target power prediction model is selected from the power prediction models based on the model weight values or the selected probabilities of the power prediction models in the optimal selection probability distribution, and the target power prediction model is called to process the predicted meteorological data to obtain power generation data of the power generation station in a target time period. The target power prediction model is selected from the power prediction models based on the model weight values or the selected probabilities of the power prediction models in the optimal selection probability distribution, and the target power prediction model is called to process the predicted meteorological data to obtain power generation data of the power generation station in a target time period. The predicted meteorological data is input into each target power prediction model to obtain power prediction data output by the target power prediction model. The target power prediction model is selected from the power prediction models based on the model weight values or the selected probabilities of the power prediction models in the optimal selection probability distribution, and the target power prediction model is called to process the predicted meteorological data to obtain power generation data of the power generation station in a target time period.

[0007] According to the power generation power prediction method provided by the application, the optimal model and the plurality of power prediction models are trained based on the following steps: A sample data set is obtained, and the sample data set includes sample meteorological data and sample power generation data of a sample power generation station in a historical time period. The plurality of power prediction models are trained based on the sample meteorological data and the sample power generation data. The plurality of power prediction models are called to perform data optimal selection labeling on the sample data set to obtain an optimal data set. The optimal model is trained based on the optimal data set.

[0008] According to the power generation power prediction method provided by the application, the optimal model and the plurality of power prediction models are trained based on the following steps: The plurality of power prediction models are called to process the sample meteorological data to obtain a plurality of predicted power generation data. The prediction errors of the plurality of power prediction models are determined based on the plurality of predicted power generation data and the corresponding sample power generation data. The optimal labels corresponding to the sample meteorological data are determined based on the plurality of prediction errors and the plurality of power prediction models. The optimal data set is constructed based on the sample meteorological data and the corresponding optimal labels.

[0009] According to the power generation power prediction method provided by the application, the optimal label corresponding to the sample meteorological data is determined based on the plurality of prediction errors and the plurality of power prediction models, and the optimal label corresponding to the sample meteorological data is determined based on the plurality of prediction errors and the plurality of power prediction models. A label model is determined from the plurality of power prediction models, and the label model is the first second preset number of power prediction models arranged in descending order of prediction error; Based on the prediction error corresponding to each label model, the model weight value of each label model is determined, and a model weight vector is determined based on the model weight value of each label model; the model weight vector contains a plurality of weight components of the power prediction model, and the weight component corresponding to each label model is the model weight value corresponding to each label model; The model weight vector is used as the optimal label corresponding to the sample meteorological data.

[0010] According to the power generation power prediction method provided by the application, the optimal label corresponding to the sample meteorological data is determined based on the plurality of prediction errors and the plurality of power prediction models, and the optimal label corresponding to the sample meteorological data is determined based on the plurality of prediction errors and the plurality of power prediction models. The smallest prediction error is determined from the plurality of prediction errors, and the power prediction model corresponding to the smallest prediction error is used as the label model; A model weight vector is determined; the model weight vector contains a plurality of weight components of the power prediction model, and the weight component corresponding to the label model is a preset positive value, and the weight component corresponding to other power prediction models is zero; The model weight vector is used as the optimal label corresponding to the sample meteorological data.

[0011] According to the power generation power prediction method provided by the application, the plurality of power prediction models are trained based on the sample meteorological data and the sample power generation power data, and the plurality of power prediction models are trained based on the sample meteorological data and the sample power generation power data. The sample data set is split to obtain a plurality of data subsets; the sample meteorological data in each data subset corresponds to the same type of meteorological feature, or corresponds to the same historical period; Based on the sample meteorological data and sample power generation power data in the plurality of data subsets, a subset power prediction model is trained; Based on the sample meteorological data and sample power generation power data in the sample data set, a full set power prediction model is trained; Based on the subset power prediction model and the full set power prediction model, the plurality of power prediction models are determined.

[0012] The application also provides a power generation power prediction device, comprising: An acquisition unit is configured to acquire prediction meteorological data of a power plant to be predicted in a target period. a preferential unit configured to input the predicted weather data into a preferential model to obtain a preferential probability distribution output by the preferential model, wherein the preferential probability distribution comprises a plurality of power prediction models and corresponding preferential conditions of the plurality of power prediction models; a prediction unit configured to filter a target power prediction model from the plurality of power prediction models based on the preferential probability distribution, and invoke the target power prediction model to process the predicted weather data to obtain power generation data of the power plant station in a target time period; The preferential model and the plurality of power prediction models are trained based on sample weather data and sample power generation data of sample power plant stations in a historical time period.

[0013] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the power generation power prediction method according to any one of the above.

[0014] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the power generation power prediction method according to any one of the above.

[0015] The power generation power prediction method, device, electronic device and storage medium provided by the application no longer passively rely on historical backtest performance to select a model, but actively judge according to future predicted weather data through a trained preferential model, and filter out a target power prediction model that is most likely to perform excellently. This strategy of dynamically selecting the optimal tool for future weather well overcomes the defects of hysteresis, non-optimality and instability in the traditional method, and can better adapt to the mutation of weather patterns, thereby greatly improving the accuracy and stability of power generation power prediction. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0017] Figure 1 is a flowchart of the power generation power prediction method provided by the application; Figure 2 is a whole architecture diagram of the power generation power prediction process provided by the application; Figure 3 is a node flowchart of the power generation power prediction process provided by the application; Figure 4 is a structural schematic diagram of a power generation power prediction device provided by the present application; Figure 5 is a structural schematic diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0018] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] Although the proportion of new energy represented by wind energy and solar energy in the power system is rapidly increasing, wind energy and solar energy have natural intermittency, volatility and uncertainty, which brings great challenges to application and popularization. Accurate new energy power prediction is a key technology to solve this problem. By accurately predicting the power generation power in a certain period of time (such as 15 minutes to 72 hours in the future), the power grid dispatching center can make reasonable dispatching plans in advance and optimize the standby capacity, thereby reducing the phenomenon of curtailment of wind and light, and thus improving the operation efficiency of the power system.

[0020] In order to realize accurate power prediction, a complex multi-stage process is currently mostly used. The core idea is to generate multiple parallel prediction results through a many-to-many combination strategy, and then select the best result through a post-selection method. Specifically, first, meteorological data from multiple sources is obtained, and then the meteorological data and historical data of the station itself are preprocessed. After that, the processed data is used to run multiple modeling algorithms in parallel, such as traditional statistical learning models, machine learning algorithms, and deep learning models, to generate multiple parallel power prediction curves. Finally, back-testing selection or similar-day selection is used to determine the final prediction result.

[0021] Among them, back-testing selection uses the actual power generation power data of the station in a certain period of time (such as the past 7 days) as a benchmark to evaluate the historical prediction performance of the above multiple models in this period of time, selects one or several models with the smallest comprehensive error index (such as root mean square error), and uses it as the optimal model to submit the prediction result of the next day.

[0022] The similar day optimization is to use the forecast weather data of the next day to find the historical dates with similar weather trends in the historical database, and then analyze which model performs best in the similar historical dates, and select the model as the optimal model to perform the prediction task of the next day.

[0023] However, it is found in the research that the above optimization scheme has inherent defects, that is: Firstly, the optimization scheme has obvious hysteresis. Backtest optimization is based on the historical performance of the model in the past period to judge its performance in the future. However, the non-stationarity of weather leads to the changeability of the applicability of the model over time, and the model with the best performance in the past may perform poorly when facing the weather pattern that may change in the future. In short, this way essentially answers the question "which model was the most accurate in the past?" without answering the real key question, that is, "which model will be the most accurate for tomorrow's weather?" It can be seen that this way ignores the inherent characteristics of the upcoming weather conditions and cannot make a truly forward-looking optimal selection.

[0024] Secondly, the optimization result has suboptimality. Although the similar day optimization considers the future weather, its optimization logic is only based on the similarity in the weather space, which is a very rough rule. The similarity of weather characteristics in the original space is not equivalent to the similarity in the hidden feature space processed by the model, so the optimal model selected by the similar day optimization may not actually bring the best prediction effect.

[0025] Finally, the prediction result has instability. Due to the above reasons, the actual prediction may be hard-switched between multiple models every day, which is likely to cause the final output prediction curve to be not continuous and stable in style and error characteristics, and the stability of the prediction is poor.

[0026] To this end, the present application provides a power generation power prediction method, which aims to solve the problems of hysteresis, suboptimality and instability existing in the existing backtest optimization and similar day optimization, by constructing an optimization model that can actively select a model according to future prediction weather conditions, realizing dynamic and forward-looking selection of power prediction models, so as to obtain more accurate and stable power generation power prediction results. Figure 1 The present application provides a flowchart of the power generation power prediction method, as shown in Figure 1 The method comprises the following steps: Step 110, obtaining prediction weather data of a power plant to be predicted in a target period; Step 120, inputting the prediction weather data into the optimization model to obtain an optimization probability distribution output by the optimization model; the optimization probability distribution contains multiple power prediction models and their corresponding optimization conditions; At step 130, the target power prediction model is selected from the plurality of power prediction models based on the preferred probability distribution, and the target power prediction model is called to process the prediction meteorological data to obtain the power generation data of the power station in the target period. The preferred model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation data of sample power stations in a historical period.

[0027] Specifically, before the power generation prediction is performed, the power station for which the power prediction needs to be performed, i.e., the to-be-predicted power station, is determined first, and the prediction meteorological data of the power station in the target period is obtained. The to-be-predicted power station can be any type of new energy power station, for example, a wind power station, a photovoltaic power station, or a hybrid new energy power station containing multiple energy forms. The target period refers to a future period for which the power prediction needs to be performed. For example, it can be 15 minutes in the future, 4 hours in the future, 72 hours in the future, etc. The prediction meteorological data is provided by a numerical weather prediction (NWP) system, and is meteorological prediction data for the geographical location of the to-be-predicted power station in the target period. In order to reduce the uncertainty of a single data source, in the embodiment of the present application, the prediction meteorological data can be obtained from multiple authoritative meteorological service agencies, for example, meteorological prediction data provided by ECMWF (European Centre for Medium-Range Weather Forecasts) and NCEP GFS (National Centers for Environmental Prediction) is obtained simultaneously. The prediction meteorological data can include wind speed, wind direction, temperature, humidity, air pressure, light intensity, total radiation, direct radiation, and other data closely related to power generation.

[0028] After the prediction meteorological data of the power plant to be predicted is determined, in the embodiment of the present application, the prediction meteorological data can be input into the optimization model to perform optimization through the model, so as to obtain the optimization probability distribution output by the optimization model. Specifically, in the entire power generation power prediction process, the optimization model plays the role of commander, and its role is to prospectively determine which one or which several of a group of preset power prediction models is likely to perform best in the target period according to the future meteorological conditions. The output of the optimization model is an optimization probability distribution, which contains multiple power prediction models and their corresponding optimization conditions. Here, the optimization condition can be represented by a probability value corresponding to each power prediction model. The higher the probability value, the greater the possibility that the optimization model considers that the power prediction model has a higher prediction accuracy in the upcoming target period; otherwise, the smaller the possibility that the optimization model considers that the power prediction model has a higher prediction accuracy in the target period.

[0029] However, it is worth noting that before applying the optimization model for optimization, the optimization model and the multiple power prediction models can be pre-trained based on the sample meteorological data and sample power generation data of the sample power plants in the historical period. Specifically, the training process of the optimization model is as follows: first, the multiple power prediction models can be trained directly based on the sample meteorological data and sample power generation data of the sample power plants in the historical period; then, the multiple power prediction models can be called to perform power backtesting on the sample meteorological data; then, the sample meteorological data is automatically labeled according to the backtesting data, and then the labeled data is used to train the initial optimization model, and finally the trained optimization model can be obtained. The initial optimization model here can be a classification model or a ranking model, such as a residual network ResNet (Residual Network), a decision tree, a gradient boosting machine, etc., which can learn to predict the optimal model combination according to the input meteorological features through training.

[0030] After obtaining the optimization probability distribution output by the optimization model, in the embodiment of the present application, the target power prediction model can be determined from the multiple power prediction models according to the optimization probability distribution. For example, the optimization probability distribution can be analyzed to select the power prediction model with the highest probability value in the optimization condition as the target power prediction model, or a probability threshold (such as 0.8) can be set, and the power prediction models with probability values exceeding the threshold are all regarded as the target power prediction model.

[0031] Figure 2 is the overall architecture diagram of the power generation power prediction process provided by the present application, as shown in Figure 2As shown, a plurality of power prediction models collectively constitute a model pool. The model pool includes various power prediction models that can perform well under different weather patterns or data conditions. Specifically, these models are trained based on sample meteorological data and sample power generation data of sample power plants in a historical period. The sample power plants can be the power plant to be predicted itself, or other power plants similar to it in geographical location, installed capacity, and equipment type. The historical period is usually a relatively long period of time, such as the past one to five years, to ensure the richness of the data samples. The sample meteorological data includes meteorological forecast data and corresponding measured meteorological data of the sample power plants in the historical period; and the sample power generation data is usually obtained from the Supervisory Control And Data Acquisition (SCADA) system of the sample power plants, and is the real power generation data recorded by the system.

[0032] To make the model pool diverse, in the embodiments of the present application, a plurality of power prediction models can be trained by combining different data sources (such as using meteorological data from ECMWF and NCEP GFS respectively), using different data preprocessing algorithms, and applying different modeling algorithms. For example, the model pool can include models based on machine learning algorithms (such as Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost)), models based on deep learning algorithms (such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer), and traditional statistical learning models (such as Autoregressive Integrated Moving Average (ARIMA)), without specific limitation in the embodiments of the present application.

[0033] After the target power prediction model is selected by the optimal probability distribution, in the embodiments of the present application, the target power prediction model can be called to process the prediction meteorological data to obtain the power generation data of the power plant in the target period. That is, the prediction meteorological data is input into the target power prediction model for calculation and processing, and finally the power generation data output by the target power prediction model can be obtained. This power generation data is a series of power generation prediction values arranged in chronological order (for example, one point every 15 minutes).

[0034] The power generation prediction method provided by the application does not passively rely on historical backtest performance to select a model, but actively judges in advance according to future predicted meteorological data by using a trained optimal selection model, and filters out a target power prediction model that is most likely to perform excellently. This strategy of dynamically selecting an optimal tool for future weather conditions well overcomes the defects of hysteresis, non-optimality and instability in traditional methods, can better adapt to mutations in weather patterns, and thus greatly improves the accuracy and stability of power generation prediction.

[0035] Based on the above embodiment, the optimal selection condition includes a model weight value or a selection probability of the corresponding power prediction model; and step 130 includes: Based on the model weight value or the selection probability of each power prediction model in the optimal selection probability distribution, a target power prediction model is filtered out from the power prediction models; and the target power prediction model is the first preset number of power prediction models in descending order of the model weight value or the selection probability. The predicted meteorological data is input into each target power prediction model to obtain power prediction data output by each target power prediction model. Based on the power prediction data output by each target power prediction model and the model weight value or the selection probability of each target power prediction model, power generation data of the power plant in the target period is determined.

[0036] Specifically, the optimal selection condition can be embodied as a model weight value or a selection probability corresponding to each power prediction model in the embodiment of the application. The model weight value can be understood as an importance score assigned to each power prediction model in the model pool by the optimal selection model. The higher the score, the greater the contribution of the optimal selection model to the model in the upcoming target period, and the higher the credibility of the prediction result. The selection probability can be understood as the possibility of the optimal selection model predicting that each power prediction model in the model pool becomes the optimal model. Generally, the sum of the selection probabilities of all models can be normalized to 1.

[0037] Correspondingly, the process of filtering out a target power prediction model from a plurality of power prediction models according to an optimal selection probability distribution and calling the target power prediction model to process predicted meteorological data to obtain power generation data of a power plant in a target period includes: Firstly, all the power prediction models in the model pool can be ranked in descending order according to the model weight values or the selection probabilities output by the selection model. Then, the first preset number of power prediction models with high rankings can be selected to form a set of target power prediction models. The first preset number is a configurable parameter, which can be set to 2, 3, 5, 8, etc. Specifically, if there are 100 power prediction models in the model pool and the first preset number is set to 5, the 5 power prediction models with the highest model weight values or the highest selection probabilities output by the selection model can be selected as the target power prediction models actually used in this prediction task.

[0038] Next, the predicted meteorological data can be input into each target power prediction model to obtain the power prediction data output by each target power prediction model. That is, after determining the set of target power prediction models composed of the first preset number of power prediction models, the predicted meteorological data of the power plant in the target time period can be input into each target power prediction model in the set in parallel or sequentially. Since the algorithms, training data, or feature engineering of each target power prediction model may be different, they will each independently perform calculations to output their respective power prediction data.

[0039] After that, the final power generation data can be determined according to the power prediction data output by each target power prediction model and the model weight values or selection probabilities of each target power prediction model. That is, in order to obtain a single and more reliable final prediction result, instead of simply selecting one from the prediction results output by multiple target power prediction models, the quantified information given by the selection model is used for weighted fusion in the embodiment of the present application. In detail, the prediction results, i.e. power prediction data, output by all target power prediction models are weighted and averaged to obtain the final prediction result, i.e. the power generation data of the power plant in the target time period.

[0040] Among them, the coefficients for weighting can be determined according to the model weight values or selection probabilities corresponding to each target power prediction model. For example, the model weight values or selection probabilities of each target power prediction model can be normalized, such as softmax calculation, then the power prediction data output by each target power prediction model is multiplied by its corresponding normalized weight, and finally all the weighted power prediction data are superimposed at each time point to obtain a final power generation data that integrates the performance of multiple models.

[0041] Compared to selecting only a single optimal model, this embodiment of the invention selects multiple top-ranked target power prediction models and weights and fuses their prediction results based on the model weight values ​​or selection probabilities given by the optimal model. This achieves a shift from single-model selection to multi-model integration. This ensemble learning strategy utilizes collective intelligence and can effectively smooth and offset the random errors and inherent biases of a single model. It avoids the risks that may arise from placing all hopes on a single model, thus making the final output power generation data not only more accurate in prediction but also more stable and robust.

[0042] Based on the above embodiments, the power generation data can be calculated using the following formula:

[0043] in, For power generation data, , and They are respectively , , Power prediction data output by the three target power prediction models. , and They are respectively , , The normalized weights corresponding to these three target power prediction models can be obtained by normalizing the model weight values ​​of these three target power prediction models.

[0044] Based on the above embodiments, the optimal model and multiple power prediction models are trained according to the following steps: Obtain the sample dataset, which includes sample meteorological data and sample power generation data of sample power plants within a historical period; Multiple power prediction models were trained based on sample meteorological data and sample power generation data. Multiple power prediction models are invoked, and the sample dataset is labeled with optimal data to obtain the optimal dataset. The optimal model is trained based on the optimal dataset.

[0045] Specifically, the training process for the optimal selection model and the power prediction model includes: First, a sample dataset needs to be obtained. This sample dataset includes sample meteorological data and sample power generation data from one or more sample power plants over a sufficiently long historical period (e.g., the past three or five years).

[0046] The sample power generation field station here can be a to-be-predicted power generation field station that needs to be predicted in the future, or other field stations with similar geographical locations, installed capacities, and device types. The sample meteorological data is multidimensional, including not only the measured meteorological data of the sample power generation field station, but also meteorological forecast data at the same period as the measured data, such as meteorological forecast data from different sources such as ECMWF and NCEP GFS.

[0047] The sample meteorological data can cover wind speed, wind direction, temperature, and light intensity. The sample power generation power data refers to real power generation power data with high resolution (such as one point every 15 minutes or 5 minutes) obtained from the SCADA system of the sample power generation field station.

[0048] Therefore, a plurality of power prediction models can be trained according to the sample meteorological data and the sample power generation power data in the sample data set. That is, after obtaining a rich sample data set, a model pool with diversity, that is, a set of a plurality of power prediction models, can be constructed in the embodiment of the present application. Diversity is the key in the construction process of the model pool, because it ensures that the model pool can cope with complex and variable weather scenarios.

[0049] Figure 3 is the node flow chart of the power generation power prediction process provided by the present application, as shown in Figure 3 To achieve diversity, a plurality of strategies can be combined to train different models in the embodiment of the present application.

[0050] For example, the model pool can include different types of algorithm models, such as XGBoost, SVM models based on machine learning, LSTM, GRU, and Transformer models based on deep learning, and traditional statistical learning models such as ARIMA.

[0051] For another example, models can be trained for different numerical weather forecast data sources, for example, one model is trained using sample meteorological data from ECMWF, and another model is trained using sample meteorological data from GFS.

[0052] For another example, even if the same algorithm and data source are used, different models can be constructed by different data preprocessing or feature combination methods.

[0053] In this way, a plurality of power prediction models with different characteristics and advantages can be trained in the embodiment of the present application.

[0054] After that, a plurality of power prediction models can be called to perform data optimization labeling on the sample data set to obtain an optimized data set. The purpose of this step is to prepare labels for the training of the optimized model. Specifically, each time unit (e.g., each day or every 4 hours) in the historical period is traversed. For each time unit, the plurality of power prediction models trained in the previous step are called, and the sample meteorological data corresponding to the time unit is used as input to perform backtesting, thereby obtaining the power prediction results output by each power prediction model.

[0055] Subsequently, the power prediction results are compared with the sample power generation data actually occurring in the time unit, and evaluation indicators such as mean absolute percentage error (MAPE), mean absolute deviation (MAE), etc. are used to evaluate the prediction error of each power prediction model. In this time unit, the power prediction model with the smallest prediction error is the optimal model, the power prediction model with the second smallest prediction error is the suboptimal model, and so on. The priority order of all power prediction models in the model pool in this time unit can be determined, and at this time, the sample meteorological data of the time unit can be used as input, and the identification (e.g., its number in the model pool) of the optimal model can be used as a label, or the identification of the optimal power prediction model in the priority order can be used as a label. According to the input and the label, a labeled data sample can be formed. Repeat this process until the entire historical period is traversed, and all generated data samples are collected to form the final optimized data set.

[0056] After that, the optimized data set can be applied to train the optimized model. Since the optimized data set essentially reveals the rule of "under what weather conditions, which model performs best", the process of training the optimized model can be understood as a process of using a machine learning or deep learning model to learn this rule. For example, a residual network (ResNet) or gradient boosting decision tree (GBDT) model can be used as the architecture of the optimized model, i.e., the initial optimized model. During training, the sample meteorological data in the optimized data set can be used as input, and the corresponding label can be used as output target. After training, the optimized model has the ability to predict which power prediction model will perform best according to the input future meteorological data.

[0057] In this embodiment of the invention, a data-driven, self-evolving training loop is achieved through systematic data acquisition, power prediction model training, optimal dataset generation, and optimal model training. This solves the technical challenge of how to objectively construct the training objective of the model selector, ensuring that the decision-making basis of the optimal model comes entirely from real historical data performance. This provides a solid and reliable foundation for making accurate and forward-looking model selections in practical applications, thereby guaranteeing the reliability and stability of the entire power prediction process.

[0058] Based on the above embodiments, multiple power prediction models are invoked to perform optimal data labeling on the sample dataset, resulting in an optimal dataset, including: Multiple power prediction models are invoked to process the sample meteorological data and obtain multiple predicted power generation data. Based on multiple predicted power generation data and corresponding sample power generation data, the prediction errors of multiple power prediction models are determined. Based on multiple prediction errors and multiple power prediction models, the optimal label corresponding to the sample meteorological data is determined. Based on the sample meteorological data and the corresponding optimal labels, an optimal dataset is constructed.

[0059] Specifically, the process of calling multiple power prediction models to perform optimal data labeling on the sample dataset to obtain the optimal dataset can include: First, multiple power prediction models can be called for backtesting and evaluation to obtain multiple predicted power generation data. That is, the entire historical sample dataset is traversed. For each time unit (e.g., one day) in the sample dataset, the corresponding meteorological data is used as input and fed into each power prediction model in the model pool. Each power prediction model calculates based on its internal algorithm and parameters and outputs a predicted power generation data for that time unit. ,in This indicates the first [number]th [unit] in the model pool at that time unit. Predicted power generation data from a power prediction model This represents the number of power prediction models in the model pool. For a given day, the model will output 96 power generation prediction values ​​at 15-minute intervals, and these 96 power generation prediction values ​​together constitute the predicted power generation data for that day.

[0060] Then, the prediction errors of the plurality of power prediction models can be determined according to the plurality of predicted power generation data and the corresponding sample power generation data. That is, after obtaining the predicted power generation data of all the power prediction models, the advantages and disadvantages of the predicted power generation data need to be quantitatively evaluated. The evaluation benchmark is the sample power generation data, that is, the actual power generation data of the sample power generation station in the time unit.

[0061] Here, specifically, each predicted power generation data obtained in the previous step can be compared with the actual sample power generation data point by point or as a whole, and the prediction error can be calculated. The prediction error can be measured by using various evaluation indexes such as MAE, MAPE and the like.

[0062] Then, the optimal selection label corresponding to the sample meteorological data and the sample power generation data can be determined according to the plurality of prediction errors and the plurality of power prediction models. That is, the plurality of prediction errors calculated in the previous step are compared and analyzed. The simplest and most direct way is to find the smallest prediction error, and determine the power prediction model corresponding to the prediction error as the optimal model of the current time unit. The identification of the optimal model is the optimal selection label corresponding to the sample meteorological data. In a more complex strategy, the optimal model can not be a single model, but a model sequence sorted according to the size of the prediction error, or a score vector representing the degree of advantage and disadvantage of each model. However, the essence is to mark the sample meteorological data of the current time unit with a label indicating “which model is suitable” based on the objective prediction error.

[0063] Finally, the optimal selection data set can be constructed according to the sample meteorological data and the corresponding optimal selection label. That is, the sample meteorological data of the current time unit is paired with the optimal selection label determined in the previous step, thereby forming a complete and labeled data sample.

[0064] By traversing all the time units in the entire historical period, the above process is repeated for each time unit. When all the time units are processed, all the generated labeled data samples are collected together to form the optimal selection data set used for training the optimal selection model.

[0065] In the embodiment of the application, the prediction error is introduced as a judgment standard to automatically and batch generate high-quality optimal selection labels for different historical meteorological scenes. This automatic and data-driven labeling method not only greatly improves the construction efficiency and objectivity of the optimal selection data set, avoids the subjectivity and high cost of manual labeling, but also ensures that the learning goal of the optimal selection model is highly consistent with the actual prediction accuracy goal, thereby greatly improving the training effect of the model.

[0066] Based on the above embodiment, the calculation of the prediction error can be realized by the following formula:

[0067] wherein, is the average absolute error of the i-th power prediction model in the model pool in any time unit when taking one day as a time unit; =96, represents the total number of power prediction values in the predicted power generation data output by the i-th power prediction model represents the j-th power prediction value in the represents the sample power generation data corresponding to the input sample meteorological data in the time unit.

[0068] Alternatively, it can be measured by the following formula:

[0069] wherein, is the average absolute percentage error of the i-th power prediction model in the model pool in any time unit when taking one day as a time unit; =96, represents the total number of power prediction values in the predicted power generation data output by the i-th power prediction model represents the j-th power prediction value in the represents the sample power generation data corresponding to the input sample meteorological data in the time unit.

[0070] Based on the above embodiment, based on a plurality of prediction errors and a plurality of power prediction models, a preferred label corresponding to sample meteorological data is determined, comprising: determining a label model from the plurality of power prediction models, the label model being the first second preset number of power prediction models in descending order of prediction error; determining a model weight value of each label model based on the prediction error corresponding to each label model, and determining a model weight vector based on the model weight value of each label model; the model weight vector contains a number of weight components of the power prediction model, and the weight component corresponding to the position of each label model is the model weight value corresponding to each label model; the model weight vector is used as the preferred label corresponding to the sample meteorological data.

[0071] ​​​​​​​​​​​​Specifically, the process of determining the optimal label corresponding to the sample meteorological data according to the plurality of prediction errors and the plurality of power prediction models specifically includes: First, a label model needs to be determined from the plurality of power prediction models.

[0072] Here, specifically, the prediction errors of all power prediction models can be sorted first, that is, arranged in order from small to large, and then the second preset number of power prediction models ranked at the front, that is, with the smallest prediction error, are selected and taken as the label model. The second preset number is a flexible configurable parameter, which can be set to 2, 3, 5, 8, etc. By selecting a small "elite model" set as the label model, the label noise caused by the occasional excellent performance of a single model can be avoided in the embodiment of the application, so that the generation process of the label is more stable and reliable.

[0073] Next, the model weight values of the label models can be determined based on the prediction errors corresponding to the label models, and a model weight vector can be determined according to the model weight values of the label models. That is, after determining the label models, a model weight value needs to be assigned to each label model to represent its relative importance on the sample meteorological data in the corresponding time unit. The determination of the model weight value should follow a basic principle, that is, the smaller the prediction error, the higher the model weight value, and vice versa, the larger the prediction error, the smaller the corresponding model weight value.

[0074] As a preferred, the model weight values of the label models are determined by the following formula in the embodiment of the application:

[0075] wherein, , and are the model weight values of the three label models 1, 2 and 3 respectively; is a coefficient for controlling the smoothness, and the concentration of the model weight value can be controlled by adjusting its size; is a normalized exponential function, , and are the mean absolute errors of the three label models 1, 2 and 3 controlled by respectively.

[0076] After the model weight values of the label models are calculated, a model weight vector needs to be constructed based on this, and the vector is a vector with a dimension equal to the number of power prediction models in the model pool (for example, if there are 100 models in total, the vector is 100-dimensional).

[0077] Specifically, this could mean that for models selected as label models from the model pool, their model weight values ​​can be filled into the weight components in the model weight vector corresponding to their positions in the model pool. For all other models in the model pool that were not selected as label models, their corresponding weight components in the model weight vector can be set to 0.

[0078] Then, the model weight vector can be used as the optimal label for the sample meteorological data. Specifically, for the current time unit, a high-dimensional, sparse model weight vector is obtained. This vector can be directly used as the optimal label for the sample meteorological data in the current time unit. Based on this sample meteorological data and the corresponding optimal label, a complete data sample can be constructed. , ,in, For the first The time unit (i.e., the first time unit) Sample meteorological data (days), That is The best label, This represents the number of time units within a historical period.

[0079] Compared to the hard-label method that selects only a single optimal model, this embodiment of the invention first selects a small set of elite models and assigns weight values ​​to them, effectively reducing the randomness and noise in the label generation process and making the labels more stable. Furthermore, the model weight vector carries richer information, not only indicating which models are good but also quantifying the degree of goodness. This transforms the training objective of the optimal model from a simple classification task to a more complex regression task, enabling it to learn more detailed and precise patterns. Consequently, the finally trained optimal model can provide a more reasonable selection criteria in practical applications.

[0080] Based on the above embodiments, and based on multiple prediction errors and multiple power prediction models, the optimal label corresponding to the sample meteorological data is determined, including: The minimum prediction error is determined from multiple prediction errors, and the power prediction model corresponding to the minimum prediction error is used as the label model. Determine the model weight vector; the model weight vector contains the number of weight components of the power prediction model, the weight component at the corresponding position of the label model is a preset positive value, and the weight component at the corresponding position of other power prediction models is zero; The model weight vector is used as the optimal label for the sample meteorological data.

[0081] Specifically, the process of determining the optimal label for the sample meteorological data based on multiple prediction errors and multiple power prediction models includes: Firstly, the minimum prediction error is determined from the plurality of prediction errors, and the power prediction model corresponding to the minimum prediction error is taken as the label model. That is, all the prediction errors are traversed to find the minimum value. Then, the power prediction model corresponding to the minimum prediction error is designated as the label model of the current time unit.

[0082] After the unique label model is determined, in the embodiment of the application, a model weight vector representing the selection result can be constructed based on the label model. The dimension of the model weight vector is consistent with the number of power prediction models in the model pool. In the vector, only the weight component at the position corresponding to the label model is a preset positive value, such as 1; the weight components at all other positions in the vector, that is, the weight components at the positions of all non-label models in the model pool, are 0.

[0083] The model weight vector generated in this way is a very sparse vector, in which only one element is a preset positive value, and all other elements are 0. Such a vector is often referred to as a One-Hot Encoding vector.

[0084] After that, the model weight vector can be taken as the preferred label corresponding to the sample meteorological data. That is, the model weight vector constructed and having only one non-zero element is taken as the preferred label corresponding to the sample meteorological data under the current time unit. Based on the sample meteorological data and the corresponding preferred label, a complete data sample can be constructed.

[0085] By traversing all the time units in the historical period, a data sample can be generated for each time unit, and finally a complete preferred data set is collected.

[0086] In the embodiment of the application, the training task of the preferred model is set as a multi-classification problem, each power prediction model corresponds to a class, and the learning goal of the preferred model is to accurately predict which “optimal class” belongs to according to the input sample meteorological data, so that the complexity of label generation is greatly reduced, and the training process of the preferred model is accelerated.

[0087] Based on the above embodiment, a plurality of power prediction models are trained based on sample meteorological data and sample power generation data, including: The sample data set is split to obtain a plurality of data subsets; the sample meteorological data in each data subset corresponds to the same meteorological feature or corresponds to the same historical period; The subset power prediction model is trained based on the sample meteorological data and the sample power generation data in the plurality of data subsets; The full-set power prediction model is trained based on the sample meteorological data and the sample power generation data in the sample data set; The multiple power prediction models are determined based on the subset power prediction model and the full-set power prediction model.

[0088] Specifically, the process of training the multiple power prediction models according to the sample meteorological data and the sample power generation data can specifically include the following steps. First, the sample data set can be split into multiple subsets to obtain multiple data subsets. Here, the sample meteorological data of all time units in the sample data set can be analyzed by using a clustering algorithm such as K-Means to aggregate sample meteorological data with similar weather characteristics together to form multiple clusters (corresponding to multiple weather types / meteorological patterns). For example, the sample meteorological data in the sample data set can be automatically clustered into several categories such as “sunny and breezy day”, “summer strong convective day”, “winter cold wave and gale day”, and “continuous rainy day”. All sample meteorological data in each category and the corresponding sample power generation data together constitute a data subset. Of course, the sample data set can also be divided by time, i.e., divided into four data subsets of spring, summer, autumn, and winter according to seasons, or divided by years to capture the impact of equipment aging or changes in operation and maintenance strategies.

[0089] Then, the subset power prediction model can be trained according to the sample meteorological data and the sample power generation data in the multiple data subsets. That is, one or more power prediction models can be trained independently for each data subset obtained in the previous step, and these models can be referred to as subset power prediction models. For example, for the data subset of “winter cold wave and gale day”, a “winter cold wave expert model” can be trained using the sample meteorological data and the sample power generation data in the data subset. Since the model only learns the data rules under a specific weather pattern, it can very finely capture the power generation characteristics under this scenario, thereby becoming an expert model in this field. By training a corresponding expert model for each data subset, a set of “each has its own strengths” subset power prediction models can be obtained.

[0090] Meanwhile, the full-set power prediction model can be trained using the sample meteorological data and the sample power generation data in the sample data set. That is, while training the expert model, at least one full-set power prediction model is also trained using the complete sample data set to ensure the generality and robustness of the model. Since the model learns all the data in the historical period, it can serve as a general model. Although the model may not perform as well as the corresponding expert model in a specific scenario, it has a wide range of applications and can provide more robust and reliable prediction results in some rare and not explicitly classified weather scenarios.

[0091] Afterwards, a plurality of power prediction models can be determined according to the subset power prediction model and the full-set power prediction model. That is, all the trained subset power prediction models (i.e., each expert model) and the full-set power prediction model (i.e., the general model) are used as power prediction models to form a model pool for the selection of an optimal model.

[0092] In the embodiment of the present application, a structured and highly diversified model pool is constructed through data set splitting, subset training and full-set training. The model pool contains both expert models with deep understanding of specific weather patterns or seasonal periods and general models with wide applicability and robust performance. The combination of such expert models and general models greatly enriches the selection space of the optimal model, ensures that there is probably one or more power prediction models highly matched with any future weather scenario in the model pool, and provides a data basis for the accurate selection of the optimal model, thereby greatly improving the accuracy and stability of the power generation prediction process.

[0093] The power generation prediction device provided by the present application is described below. The power generation prediction device described below can be referred to in conjunction with the power generation prediction method described above.

[0094] Figure 4 is a structural schematic diagram of the power generation prediction device provided by the present application, as shown in Figure 4 The device comprises: An acquisition unit 410 is configured to acquire prediction meteorological data of a power plant station to be predicted in a target period. An optimal selection unit 420 is configured to input the prediction meteorological data into an optimal selection model to obtain an optimal probability distribution output by the optimal selection model. The optimal probability distribution contains a plurality of power prediction models and corresponding optimal selection conditions thereof. A prediction unit 430 is configured to filter a target power prediction model from the plurality of power prediction models based on the optimal probability distribution, and call the target power prediction model to process the prediction meteorological data to obtain power generation data of the power plant station in the target period. The optimal selection model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation data of sample power plant stations in a historical period.

[0095] The power generation power prediction device provided by the application is no longer passively dependent on historical backtest performance to select a model, but actively judges and filters out the target power prediction model most likely to perform excellently according to future prediction meteorological data through the trained optimal selection model. This strategy of dynamically selecting the optimal tool for future weather well overcomes the defects of hysteresis, non-optimality and instability in the traditional method, can better adapt to the mutation of weather patterns, and greatly improves the accuracy and stability of power generation power prediction.

[0096] Based on the above embodiment, the selected optimal case includes the model weight value or the selected probability of the corresponding power prediction model; the prediction unit 430 is used to: Based on the model weight value or the selected probability of each power prediction model in the optimal selection probability distribution, the target power prediction model is filtered out from the power prediction models; the target power prediction model is the first preset number of power prediction models in the order from high to low of the model weight value or the selected probability; The prediction meteorological data is input into each target power prediction model to obtain power prediction data output by each target power prediction model; Based on the power prediction data output by each target power prediction model and the model weight value or the selected probability of each target power prediction model, the power generation power data of the power station in the target period is determined.

[0097] Based on the above embodiment, the device further includes a training unit for: Obtaining a sample data set, the sample data set including sample meteorological data and sample power generation power data of a sample power station in a historical period; Based on the sample meteorological data and the sample power generation power data, the plurality of power prediction models are trained to obtain; Calling the plurality of power prediction models, data optimal labeling is performed on the sample data set to obtain an optimal data set; Based on the optimal data set, the optimal model is trained to obtain.

[0098] Based on the above embodiment, the training unit is used to: Calling the plurality of power prediction models, the sample meteorological data is processed to obtain a plurality of predicted power generation power data; Based on the plurality of predicted power generation power data and the corresponding sample power generation power data, the prediction error of the plurality of power prediction models is determined; Based on the plurality of prediction errors and the plurality of power prediction models, the optimal label corresponding to the sample meteorological data is determined; Based on the sample meteorological data and the corresponding optimal selection label, the optimal selection data set is constructed.

[0099] Based on the above embodiment, the training unit is configured to: Determine a label model from the plurality of power prediction models, the label model being the first second preset number of power prediction models in descending order of prediction error; Determine a model weight value of each label model based on the prediction error corresponding to the label model, and determine a model weight vector based on the model weight values of the label models; the model weight vector includes a number of weight components of the power prediction models, and the weight component at the position corresponding to the label model is the model weight value corresponding to the label model; The model weight vector is used as the optimal selection label corresponding to the sample meteorological data.

[0100] Based on the above embodiment, the training unit is configured to: Determine the minimum prediction error from the plurality of prediction errors, and use the power prediction model corresponding to the minimum prediction error as the label model; Determine a model weight vector; the model weight vector includes a number of weight components of the power prediction models, the weight component at the position corresponding to the label model is a preset positive value, and the weight component at the position corresponding to other power prediction models is zero; The model weight vector is used as the optimal selection label corresponding to the sample meteorological data.

[0101] Based on the above embodiment, the training unit is configured to: Split the sample data set to obtain a plurality of data subsets; the sample meteorological data in each data subset corresponds to the same type of meteorological feature, or corresponds to the same historical period; Based on the sample meteorological data and the sample power generation data in the plurality of data subsets, a subset power prediction model is trained; Based on the sample meteorological data and the sample power generation data in the sample data set, a full set power prediction model is trained; Determine the plurality of power prediction models based on the subset power prediction model and the full set power prediction model.

[0102] Figure 5 An example of an entity structure diagram of an electronic device is shown as Figure 5As shown, the electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 complete mutual communication through the communications bus 540. The processor 510 can invoke a logic instruction in the memory 530 to execute a power generation power prediction method, which includes: obtaining predicted meteorological data of a power generation station to be predicted in a target period; inputting the predicted meteorological data into a selection model to obtain a selection probability distribution output by the selection model; the selection probability distribution contains a plurality of power prediction models and their corresponding selected conditions; based on the selection probability distribution, a target power prediction model is selected from the plurality of power prediction models, and the target power prediction model is called to process the predicted meteorological data to obtain power generation power data of the power generation station in the target period; wherein the selection model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation power data of sample power generation stations in a historical period.

[0103] In addition, the logic instruction in the memory 530 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0104] In another aspect, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions that, when executed by a computer, enable the computer to perform the power generation prediction method provided by any of the above methods, the method comprising: obtaining predicted meteorological data of a power generation station to be predicted in a target time period; inputting the predicted meteorological data into a selection model to obtain a selection probability distribution output by the selection model; the selection probability distribution comprising a plurality of power prediction models and corresponding selection conditions thereof; based on the selection probability distribution, screening a target power prediction model from the plurality of power prediction models, and calling the target power prediction model to process the predicted meteorological data to obtain power generation data of the power generation station in the target time period; wherein the selection model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation data of sample power generation stations in a historical time period.

[0105] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a power generation prediction method provided by any of the above methods, the method comprising: obtaining predicted meteorological data of a power generation station to be predicted in a target time period; inputting the predicted meteorological data into a selection model to obtain a selection probability distribution output by the selection model; the selection probability distribution comprising a plurality of power prediction models and corresponding selection conditions thereof; based on the selection probability distribution, screening a target power prediction model from the plurality of power prediction models, and calling the target power prediction model to process the predicted meteorological data to obtain power generation data of the power generation station in the target time period; wherein the selection model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation data of sample power generation stations in a historical time period.

[0106] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0107] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in the various embodiments or some parts of the embodiments.

[0108] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of power generation forecast, characterized by, include: Obtain the predicted meteorological data for the power plants to be predicted within the target time period; The predicted meteorological data is input into the selection model to obtain the selection probability distribution output by the selection model; the selection probability distribution includes multiple power prediction models and their corresponding selection cases; Based on the optimal probability distribution, a target power prediction model is selected from the multiple power prediction models, and the target power prediction model is called to process the predicted meteorological data to obtain the power generation data of the power plant in the target time period. The optimal model and the multiple power prediction models are trained based on sample meteorological data and sample power generation data of sample power plants within historical time periods.

2. The power generation power prediction method according to claim 1, characterized by, The selection criteria include the model weight value or the probability of selection for the corresponding power prediction model; Based on the optimal probability distribution, a target power prediction model is selected from the multiple power prediction models, and the target power prediction model is used to process the predicted meteorological data to obtain the power generation data of the power plant during the target time period, including: Based on the model weight value or selection probability of each power prediction model in the optimal probability distribution, a target power prediction model is selected from the power prediction models; the target power prediction model is the first preset number of power prediction models when arranged in descending order of model weight value or selection probability. The predicted meteorological data is input into each target power prediction model to obtain the power prediction data output by each target power prediction model; Based on the power prediction data output by each target power prediction model, and the model weight value or selection probability of each target power prediction model, the power generation data of the power plant in the target time period is determined.

3. The power generation power prediction method according to claim 1, characterized by, The optimal model and the multiple power prediction models are trained based on the following steps: Obtain a sample dataset, which includes sample meteorological data and sample power generation data of sample power plants within a historical period; Based on the sample meteorological data and the sample power generation data, the multiple power prediction models are trained. The multiple power prediction models are invoked to perform optimal data labeling on the sample dataset, resulting in an optimal dataset. The optimal model is trained based on the aforementioned optimal dataset.

4. The power generation power prediction method according to claim 3, characterized by, The process of calling the multiple power prediction models and performing optimal data labeling on the sample dataset to obtain the optimal dataset includes: The multiple power prediction models are invoked to process the sample meteorological data, resulting in multiple predicted power generation data. Based on multiple predicted power generation data and corresponding sample power generation data, the prediction error of the multiple power prediction models is determined; Based on multiple prediction errors and the multiple power prediction models, the optimal label corresponding to the sample meteorological data is determined; Based on the sample meteorological data and the corresponding preferred labels, the preferred dataset is constructed.

5. The power generation power prediction method according to claim 4, characterized by, The process of determining the optimal label corresponding to the sample meteorological data based on multiple prediction errors and multiple power prediction models includes: determining a label model from the plurality of power prediction models, the label model being a first preset number of power prediction models in descending order of prediction errors; determining model weight values of the label models based on the prediction errors corresponding to the label models, and determining a model weight vector based on the model weight values of the label models; the model weight vector including a number of weight components of the power prediction models, and a weight component at a position corresponding to a label model being a model weight value corresponding to the label model; using the model weight vector as a preferred label corresponding to the sample meteorological data.

6. The power generation power prediction method according to claim 4, characterized by, The preferred label corresponding to the sample meteorological data is determined based on the plurality of prediction errors and the plurality of power prediction models, including: determining a minimum prediction error from the plurality of prediction errors, and using a power prediction model corresponding to the minimum prediction error as a label model; determining a model weight vector; the model weight vector including a number of weight components of the power prediction models, a weight component at a position corresponding to the label model being a preset positive value, and weight components at positions corresponding to other power prediction models being zero; using the model weight vector as a preferred label corresponding to the sample meteorological data.

7. The power generation forecast method according to any one of claims 3 to 6, characterized by, The plurality of power prediction models are trained based on the sample meteorological data and the sample power generation data, including: splitting the sample data set to obtain a plurality of data subsets; sample meteorological data in each data subset corresponding to a same type of meteorological feature or a same historical time period; training a subset power prediction model based on sample meteorological data and sample power generation data in the plurality of data subsets; training a full-set power prediction model based on sample meteorological data and sample power generation data in the sample data set; determining the plurality of power prediction models based on the subset power prediction model and the full-set power prediction model.

8. A power generation power prediction device characterized by comprising: including: an acquisition unit configured to acquire prediction meteorological data of a power plant to be predicted in a target time period; a preferred unit configured to input the prediction meteorological data into a preferred model to obtain a preferred probability distribution output by the preferred model; the preferred probability distribution including a plurality of power prediction models and corresponding preferred conditions thereof; a prediction unit configured to select a target power prediction model from the plurality of power prediction models based on the preferred probability distribution, and call the target power prediction model to process the prediction meteorological data to obtain power generation data of the power plant in the target time period; wherein the preferred model and the plurality of power prediction models are trained based on sample meteorological data and sample power generation data of sample power plants in a historical time period.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to implement the power generation power prediction method of any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the power generation power prediction method of any one of claims 1 to 7.

Images