A probabilistic prediction method for power generation of distributed photovoltaic sites in a region

Abstract

The present invention discloses a probabilistic prediction method for the power generation of distributed photovoltaic sites in a region, and the specific steps are as follows: step 1: establishing a prediction model; step 2: checking whether the prediction model needs to be updated; step 3: training each prediction model using training data; step 4: finding the optimal weight of each prediction model; step 5: deriving the error distribution from the ensemble model; step 6: predicting deterministic photovoltaic power generation through the ensemble model; step 7: generating PI from the error distribution and the deterministic prediction; the method develops a universal model and method to predict the distributed photovoltaic power generation in the region, and uses three data-driven regression models, namely K-Means, Naive Bayes classifier and neural network, to perform ensemble prediction to process the power generation of multiple PVs geographically distributed in a specific area. Simulation shows that when the confidence level is 95%, the mean square error between the power generation prediction value and the actual value of the proposed method is about 0.148kW, which verifies the robustness of the method.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation, and in particular to a method for probabilistically predicting power generation of distributed photovoltaic sites within a region. Background Art

[0002] Over the past decade, the global photovoltaic market has grown rapidly. According to the International Energy Agency, global photovoltaic production capacity has exceeded 500 GWp. Distributed photovoltaic power generation refers specifically to photovoltaic power generation facilities built near user sites, operating in a manner characterized by self-consumption at the user's site, supplying excess electricity to the grid, and balancing regulation within the distribution system. Photovoltaic power generation depends on many meteorological variables, such as solar irradiance, air temperature, cloudiness, wind speed, and relative humidity. Forecasting distributed photovoltaic power generation is a challenge, especially in the context of multi-step applications, large databases, noise measurements, and multiple input and output observations. Generally speaking, the factors that affect the accuracy of photovoltaic power generation forecasts include: time range and temporal resolution, weather conditions, geographic location, and data availability and quality.

[0003] Existing distributed photovoltaic power generation forecasting methods can be divided into two main categories. One category improves forecast accuracy by building spatiotemporal models to extract and utilize spatiotemporal data from multiple photovoltaic sources. For example, photovoltaic power generation forecasting methods based on deep learning frameworks can generate photovoltaic power generation forecasts for multiple regions and areas. A cloud motion vector-based method is an established method for forecasting distributed photovoltaic power generation covering a specific area. Tracking-based distributed photovoltaic power generation forecasts have a mean absolute percentage error of 4.23% at a 30-minute forecast horizon, with the error increasing with longer forecast horizons. Numerical weather prediction models are widely used to predict atmospheric conditions up to 15 days out and do not require any historical data. These methods offer good accuracy but rely heavily on the stability of weather conditions. Furthermore, the implementation of these physical models is generally difficult, requiring known parameters and expensive equipment to construct them. These limitations make them unavailable in many regions. Furthermore, for most available numerical weather prediction models, the value of data available within a few hours of the forecast is limited. The other category is photovoltaic power generation forecasting methods driven by historical data. For example, photovoltaic power generation prediction methods based on Delta, Bayesian, mean-variance estimation and bootstrap techniques respectively develop prediction intervals through data-driven development for different prediction objectives.

[0004] However, existing methods for modeling multiple distributed PV regions are limited. Most studies focus on predictions at a single location, while research on regional models is rare. Currently, there is limited research on predicting distributed PV power generation, primarily focusing on regional cumulative PV power generation rather than individual PV generation. Furthermore, existing prediction models are overly specific and limited to specific, idealized regions. Therefore, there is an urgent need to develop universal models and methods for predicting distributed PV power generation within a region. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for probabilistically predicting the power generation of distributed photovoltaic sites in a region, so as to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A probabilistic prediction method for power generation of distributed photovoltaic sites in a region, comprising the following steps:

[0008] Step 1: Establish the prediction model, that is, establish the integrated model of the 1st to ith PV nodes, where the PV nodes are constant voltage nodes;

[0009] Step 2: Check whether the prediction model needs to be updated;

[0010] Step 3: Use the training data to train each prediction model;

[0011] Step 4: Find the optimal weights for each prediction model;

[0012] Step 5: Derive the error distribution from the ensemble model and determine the boundaries of the PI nodes based on the data of the PV nodes. The PI nodes are constant current nodes.

[0013] Step 6: Predict the deterministic PV power generation through the ensemble model. After developing the error distribution during the model training process, predict the deterministic PV node power generation for the next 24 hours.

[0014] Step 7: Generate a PI node from the error distribution and the deterministic prediction, where the PI node includes an upper limit and a lower limit.

[0015] Preferably, the prediction model in step 1 includes two types: a single prediction model and a multiple prediction model:

[0016] The single prediction model consists of each PV power generation site, represented as node, node, …, node. Based on historical power generation and weather data, the single prediction model is used to predict the power generation from node to node;

[0017] The multi-prediction model is used to predict the power generation of the node. The input of the multi-prediction model is the predicted power generation of each photovoltaic power generation by the single prediction model, historical data and weather data;

[0018] Three data-driven regression models, namely K-Means, Naive Bayes classifier and neural network, are used for ensemble prediction to handle multiple PVs geographically distributed in a specific area. In the proposed method, an ensemble model is arranged for each PV and a prediction model is constructed. Each data-driven regression model is trained individually to configure the best parameters based on past data.

[0019] Preferably, the specific steps in step 2 are as follows:

[0020] 1) Check whether the parameters of the ensemble prediction model are up to date;

[0021] 2) If the trained model with established parameters does not contain the latest observation data, the model is retrained to update the parameters of each prediction method;

[0022] 3) During the training process, the parameters of the ensemble prediction method are determined using past data;

[0023] 4) After the training process is completed, the parameters of the prediction method remain unchanged until a new training process is performed;

[0024] 5) Model parameters must be updated regularly.

[0025] Preferably, the specific method in step three is as follows:

[0026] Train the K-Means model, the Naive Bayes classifier model, and the NN model separately:

[0027] Model building and forecasting are divided into two groups: historical data and forecast data:

[0028] Historical data, i.e., training and validation data, contain predictor variables and targets. Predictor variables include timestamp, temperature, and weather conditions, and the target is photovoltaic power generation.

[0029] The forecast data only contains the forecast variables. In the forecast data, weather information is obtained through weather forecasts.

[0030] Historical data contain at least one year of collected data to capture seasonal characteristics;

[0031] The validation data is a set of historical data of arbitrary length, which is used to determine the optimal weights of the ensemble prediction model and to construct the error distribution, thereby generating a prediction interval. The cause of the error can be explained based on the validation data and the error distribution.

[0032] The NN model consists of four hidden layers with 25, 25, 25, and 20 units, respectively. The scaled conjugate gradient is used as the training function, and the prediction results of each time step are averaged three times to mitigate the impact of random initial values.

[0033] The K-Means method is used to classify the training data. The classified training data is used as the input of the Naive Bayes classifier. The process of combining K-Means and Naive Bayes classifiers to predict PV power generation is as follows:

[0034] Step 1 : K-Means classifies the observed PV power generation records into different clusters, and then classifies the predictors such as temperature and weather conditions corresponding to each timestamp into each cluster;

[0035] Step 2 :Training a naive Bayes classifier model using categorical observations and a predicted kernel distribution function;

[0036] Step 3 : The trained Naive Bayes classifier classifies the unknown predictor variables into the test data for each cluster;

[0037] Step 4 : Step 1 The centroid of each cluster identified in is the predicted PV power generation value.

[0038] Preferably, the specific method in step 4 is as follows:

[0039] First, the optimal coefficients of the ensemble model consisting of two training models, the naive Bayes classifier and the neural network, are determined. For the naive Bayes classifier model, an optimal K is determined, which represents the number of groups that need to be generated. The NN model learns the weight of each neuron. The two prediction models are combined with the weights to construct the ensemble prediction model.

[0040] Preferably, the specific method in step 5 is as follows:

[0041] Based on the optimal coefficients obtained in step 4, the trained ensemble prediction model is then used to predict future PV node power generation. The boundaries of the PI nodes can also be further obtained, and the absolute error set at a specific time is obtained by comparing the predicted data and the observed data throughout the entire history.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This method develops a universal model and method to predict distributed photovoltaic power generation in a region. It uses three data-driven regression models, K-Means, Naive Bayes classifier and neural network, to perform ensemble prediction to process the power generation of multiple PVs geographically distributed in a specific area. Simulation shows that when the confidence level is 95%, the mean square error between the power generation prediction value and the actual value of the proposed method is about 0.148kW, which verifies the robustness of this method and provides an effective and practical mechanism for predicting distributed photovoltaic power generation in a region. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a configuration diagram of single and multiple PV forecast models;

[0045] Figure 2 It is the configuration diagram of the ensemble prediction model;

[0046] Figure 3 It is a flow chart of ensemble prediction;

[0047] Figure 4 It is a data configuration diagram;

[0048] Figure 5 It is the prediction effect diagram with a confidence level of 90%;

[0049] Figure 6 This is the prediction effect diagram with a confidence level of 95%. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solution in conjunction with the embodiments of the present invention. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention. Example

[0051] See also Figure 1-6 The present invention provides a technical solution: a method for probabilistically predicting power generation of distributed photovoltaic sites within a region, the specific steps of which are as follows:

[0052] Step 1: Establish the prediction model, that is, establish the integrated model of the 1st to ith PV nodes, where the PV nodes are constant voltage nodes;

[0053] Step 2: Check whether the prediction model needs to be updated;

[0054] Step 3: Use the training data to train each prediction model;

[0055] Step 4: Find the optimal weights for each prediction model;

[0056] Step 5: Derive the error distribution from the ensemble model and determine the boundaries of the PI nodes based on the data of the PV nodes. The PI nodes are constant current nodes.

[0057] Step 6: Predict the deterministic PV power generation through the ensemble model. After developing the error distribution during the model training process, predict the deterministic PV node power generation for the next 24 hours.

[0058] Step 7: Generate a PI node from the error distribution and the deterministic prediction, where the PI node includes an upper limit and a lower limit.

[0059] Specifically, the single prediction model consists of each PV power generation site and is expressed as node, node,…, Node, based on historical power generation and weather data, a single prediction model is used to predict Node to The power generation of the node;

[0060] Multiple prediction models are used for prediction The power generation of the node, the input of the multi-prediction model is the predicted power generation of each PV power generation by the single prediction model, historical data and weather data;

[0061] Three data-driven regression models, namely K-Means, Naive Bayes classifier and neural network, are used for ensemble prediction to handle multiple PVs geographically distributed in a specific area. In the proposed method, an ensemble model is arranged for each PV and a prediction model is constructed. Each data-driven regression model is trained individually to configure the optimal parameters based on past data.

[0062] Specifically, the specific steps in step 2 are as follows:

[0063] 1) Check whether the parameters of the ensemble prediction model are up to date;

[0064] 2) If the trained model with established parameters does not contain the latest observation data, the model is retrained to update the parameters of each prediction method;

[0065] 3) During the training process, the parameters of the ensemble prediction method are determined using past data;

[0066] 4) After the training process is completed, the parameters of the prediction method remain unchanged until a new training process is performed;

[0067] 5) Model parameters must be updated regularly.

[0068] Specifically, the specific method in step three is as follows:

[0069] Train the K-Means model, the Naive Bayes classifier model, and the NN model separately:

[0070] Model building and forecasting are divided into two groups: historical data and forecast data:

[0071] Historical data, i.e., training and validation data, contain predictor variables and targets. Predictor variables include timestamp, temperature, and weather conditions, and the target is photovoltaic power generation.

[0072] The forecast data only contains the forecast variables. In the forecast data, weather information is obtained through weather forecasts.

[0073] Historical data contain at least one year of collected data to capture seasonal characteristics;

[0074] The validation data is a set of historical data of arbitrary length, which is used to determine the optimal weights of the ensemble prediction model and to construct the error distribution, thereby generating a prediction interval. The cause of the error can be explained based on the validation data and the error distribution.

[0075] The NN model consists of four hidden layers with 25, 25, 25, and 20 units, respectively. The scaled conjugate gradient is used as the training function, and the prediction results of each time step are averaged three times to mitigate the impact of random initial values.

[0076] The K-Means method is used to classify the training data. The classified training data is used as the input of the Naive Bayes classifier. The process of combining K-Means and Naive Bayes classifiers to predict PV power generation is as follows:

[0077] Step 1 : K-Means classifies the observed PV power generation records into different clusters, and then classifies the predictors such as temperature and weather conditions corresponding to each timestamp into each cluster;

[0078] Step 2 :Training a naive Bayes classifier model using categorical observations and a predicted kernel distribution function;

[0079] Step 3 : The trained Naive Bayes classifier classifies the unknown predictor variables into the test data for each cluster;

[0080] Step 4 : Step 1 The centroid of each cluster identified in is the predicted PV power generation value.

[0081] Specifically, the specific method in step 4 is as follows:

[0082] First, the optimal coefficients of the ensemble model consisting of two training models, the naive Bayes classifier and the neural network, are determined. For the naive Bayes classifier model, an optimal K is determined, which represents the number of groups that need to be generated. The NN model learns the weight of each neuron. The two prediction models are combined with the weights to construct the ensemble prediction model.

[0083] Specifically, the specific method in step 5 is as follows:

[0084] Based on the optimal coefficients obtained in step 4, the trained ensemble prediction model is then used to predict future PV node power generation. The boundaries of the PI nodes can also be further obtained, and the absolute error set at a specific time is obtained by comparing the predicted data and the observed data throughout the entire history.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A probabilistic prediction method for power generation of distributed photovoltaic sites within a region, characterized by: The specific steps are as follows: Step 1: Establish the prediction model, that is, establish the 1st to ith set model of the PV node. The PV node is a constant voltage node. The specific method is as follows: Prediction models include single prediction models and multi-prediction models: The single prediction model consists of each photovoltaic power generation site and is expressed as node, node,…, Node, based on historical power generation and weather data, a single prediction model is used to predict Node to The power generation of the node; Multiple prediction models are used for prediction The power generation of the node, the input of the multi-prediction model is the predicted power generation of each PV power generation by the single prediction model, historical data and weather data; The three data-driven regression models, K-Means, Naive Bayes classifier and neural network, are used for ensemble prediction to handle multiple geographically distributed areas. Node,In the proposed method, an ensemble model is arranged for each PV and a prediction model is constructed, and each data-driven regression model is trained individually to configure the best parameters based on past data; Step 2: Check whether the prediction model needs to be updated; Step 3: Use the training data to train each prediction model. The specific method is as follows: The K-Means method is used to classify the training data. The classified training data is used as the input of the Naive Bayes classifier. The process of combining K-Means and Naive Bayes classifiers to predict PV power generation is as follows: Step 1 : K-Means classifies the observed PV power generation records into different clusters, and then classifies the temperature and weather condition predictors corresponding to each timestamp into each cluster; Step 2 :Training a naive Bayes classifier model using categorical observations and a predicted kernel distribution function; Step 3 : The trained Naive Bayes classifier classifies the unknown predictor variables into the test data for each cluster; Step 4 : Step 1 The centroid of each cluster identified in is the predicted PV power generation value; Step 4: Find the optimal weights for each prediction model. The specific method is as follows: First, we determine the optimal weights of the ensemble model consisting of two training models: the Naive Bayes classifier and the neural network. For the Naive Bayes classifier model, we determine an optimal K, which represents the number of groups to be generated. The NN model learns the weights of each neuron. We combine these two prediction models with the weights to construct the ensemble prediction model. Step 5: Derive the error distribution from the ensemble model and determine the boundaries of the PI nodes based on the data of the PV nodes. The PI nodes are constant current nodes. Step 6: Predict the deterministic PV power generation through the ensemble model. After developing the error distribution during the model training process, predict the deterministic PV node power generation for the next 24 hours. Step 7: Generate a PI node from the error distribution and the deterministic prediction, where the PI node includes an upper limit and a lower limit.

2. The method for predicting the power generation of distributed photovoltaic sites in a region according to claim 1, characterized in that: The specific steps in step 2 are as follows: 1) Check whether the parameters of the ensemble prediction model are up to date; 2) If the trained model with established parameters does not contain the latest observation data, the model is retrained to update the parameters of each prediction method; 3) During the training process, the parameters of the ensemble prediction method are determined using past data; 4) After the training process is completed, the parameters of the prediction method remain unchanged until a new training process is performed; 5) Model parameters must be updated regularly.

3. The method for predicting the power generation of distributed photovoltaic sites in a region according to claim 1, characterized in that: The specific method in step three is as follows: Train the K-Means model, the Naive Bayes classifier model, and the NN model separately: Data model building and prediction are divided into two groups: historical data and forecast data: Historical data, i.e., training and validation data, contain predictor variables and targets. Predictor variables include timestamp, temperature, and weather conditions, and the target is photovoltaic power generation. The forecast data only contains the forecast variables. In the forecast data, weather information is obtained through weather forecasts. Historical data contain at least one year of collected data to capture seasonal characteristics; The validation data is a set of historical data of arbitrary length, which is used to determine the optimal weights of the ensemble prediction model and to construct the error distribution, thereby generating a prediction interval. The cause of the error can be explained based on the validation data and the error distribution. The NN model consists of four hidden layers with 25, 25, 25, and 20 units, respectively. The scaled conjugate gradient is used as the training function, and the prediction results of each time step are averaged three times to mitigate the impact of random initial values.

4. The method for predicting the power generation of distributed photovoltaic sites in a region according to claim 1, characterized in that: The specific method in step five is as follows: Based on the optimal weights obtained in step 4, the trained ensemble prediction model is then used to predict the future PV node power generation, and the boundaries of the PI nodes are further obtained. The absolute error set at a specific time is obtained by comparing the predicted data and the observed data throughout the entire history.

Images