Photovoltaic power generation interval prediction method fusing multiple targets

Abstract

The invention discloses a photovoltaic power generation interval prediction method fusing multiple targets. Comprising model construction, the model construction comprises the construction of a dynamic interval optimization loss function, and the construction of the dynamic interval optimization loss function comprises the steps that a Transform model architecture is adopted as a basic model architecture, an interval prediction loss function module is designed on the basis of a traditional regression task, and an interval prediction loss function module is designed on the basis of the traditional regression task; a multi-objective loss function fusing point prediction and interval prediction and a penalty function are provided, the optimization objective is to minimize the average prediction interval width and maximize the interval coverage rate at the same time, and long-time-sequence dynamic interval prediction and point prediction are achieved. According to the photovoltaic power generation interval prediction method fusing multiple targets, a narrow prediction interval is generated while the high coverage rate is ensured, and the defects of a traditional prediction method in the aspects of accuracy and dynamic adaptive capacity are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power generation interval prediction method integrating multiple objectives. Background Art

[0002] Currently, the existing point forecast of photovoltaic power generation cannot reflect the uncertainty of power generation. In reality, photovoltaic power generation is greatly affected by light conditions and meteorological factors, which lead to high uncertainty and volatility of photovoltaic power generation. Therefore, simply providing a specific power generation forecast value is often not enough to support the decision-making needs of the power grid.

[0003] However, the application of existing deep learning models in interval prediction has not been fully developed, and many methods still remain at the output of a single prediction value, failing to effectively construct a dynamic and adjustable prediction interval. Although there are some statistical and machine learning-based methods for interval prediction of photovoltaic power generation, there are still some shortcomings, such as the traditional prediction interval construction method is too simple and unreasonable, only considering the results of single-point prediction, while the actual needs require long-term prediction results, without considering the accuracy of point prediction, but only considering the effect of interval prediction, poor interval evaluation effect, non-differentiable loss function, lack of dynamic adjustment ability, inability to perform multi-step prediction, etc., and often fail to fully utilize the characteristics of time series data.

[0003] Therefore, for the interval prediction of photovoltaic power generation, a new method is urgently needed that can combine the advantages of deep learning, comprehensively consider the various factors that affect the power generation, and dynamically adjust the prediction interval. This is a major challenge facing the current technical field. Summary of the invention

[0004] The purpose of the present invention is to provide a photovoltaic power generation interval prediction method integrating multiple objectives in view of the shortcomings of the above-mentioned prior art. The photovoltaic power generation interval prediction method integrating multiple objectives generates a narrow prediction interval while ensuring a high coverage rate, thus solving the shortcomings of traditional prediction methods in terms of accuracy and dynamic adaptability.

[0004] The technical solution of the photovoltaic power generation interval prediction method integrating multiple objectives of the present invention includes model construction, which includes the construction of a dynamic interval optimization loss function. The construction of the dynamic interval optimization loss function includes adopting a Transformer model architecture as a basic model architecture, designing an interval prediction loss function module on the basis of traditional regression tasks, proposing a multi-objective loss function integrating point prediction and interval prediction and a penalty function, wherein the optimization goal is to minimize the average prediction interval width while maximizing the interval coverage, thereby realizing long-time series dynamic interval prediction and point prediction.

[0005] It also includes feature construction, which includes using a sliding window to construct statistical features of active power and lag features to form a time series data set.

[0006] Compared with the prior art, the photovoltaic power generation interval prediction method integrating multiple objectives of the present invention has the following beneficial effects: First, the algorithm can output long time series. The designed model is improved based on the Transformer architecture, which not only takes advantage of the significant effect of the attention mechanism in long time series prediction, but also ensures that a reasonable prediction interval can be generated. This feature enables the system to fully capture and utilize historical data in a complex power generation environment to meet the needs of long-term prediction. This is of great significance for power grid dispatching and resource planning.

[0007] Secondly, the present invention designs an interval prediction loss function module, and proposes a multi-objective loss function that integrates point prediction and interval prediction, as well as a penalty function. The optimization target comprehensively considers the coverage and average width of the prediction interval. By accurately controlling the range of the prediction interval, it not only ensures high coverage, but also minimizes the interval width, which can effectively improve the accuracy and operability of the prediction, and the model has a certain robustness. Reasonable prediction intervals and point prediction accuracies can be obtained in different application scenarios, and the integrated output of interval prediction and point prediction ensures that the model can maintain the accuracy of the prediction while pursuing the generation of reasonable prediction intervals.

[0008] In addition, the system of the present invention has the function of updating results in real time, and can automatically adjust the prediction model according to the latest meteorological data. By inputting the confidence of a given target, the corresponding prediction interval and prediction value can be obtained, which enables the model to achieve an ideal balance and performance in different scenarios. Its dynamic adaptability ensures that the prediction results are always up-to-date and accurate, thereby enhancing the ability to respond to sudden weather changes, improving the reliability of the overall prediction, and achieving self-adaptation, self-regulation, and accurate prediction.

[0009] The present invention not only improves the accuracy and scientificity of photovoltaic power generation prediction, but also provides more reliable decision-making support for the dispatch and management of power grids, helps achieve the goals of efficient utilization of renewable energy and green and low-carbon development, and promotes the development of power systems in a more intelligent and sustainable direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flow chart of an embodiment of a photovoltaic power generation interval prediction method based on a dynamic interval optimization loss function of the present invention; Figure 2 It is the composition of the loss function module. DETAILED DESCRIPTION

[0011] In order to better understand the present invention, the technical solution of the present invention is now clearly and completely described through embodiments in combination with the accompanying drawings. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form a new embodiment. It will be understood that other embodiments can be conceived and made without departing from the scope or spirit of the present disclosure. Therefore, the following detailed description should not be considered to have a limiting meaning. The photovoltaic power generation interval prediction method integrating multiple objectives of the present invention comprises the following steps:

[0012] S1. Collect historical photovoltaic power data, meteorological data and equipment operation status data, and obtain photovoltaic power data and meteorological data from the energy management system; for equipment operation status data, delete the data that does not generate electricity at night, perform outlier detection and missing value processing, and use linear interpolation to fill the missing values ​​to keep the time series data continuous and smooth in time; among them, meteorological data include: light intensity, scattered radiation, ambient temperature, relative humidity, air pressure, wind speed, wind direction; at the same time, obtain weather forecast data with a granularity of 15 minutes.

[0013] S2, feature construction, based on the corresponding historical data collected in step S1, use the sliding window to construct statistical features of active power and lag features to form a time series data set, which helps the model capture long-term and short-term change trends.

[0014] S3, model construction, including the construction of "Dynamic Interval Optimization Loss" (DIO-Loss), using the Transformer model architecture as the basic model architecture, designing the interval prediction loss function module based on the traditional regression task, proposing a multi-objective loss function that integrates point prediction and interval prediction and a penalty function, the optimization goal is to minimize the average prediction interval width while maximizing the interval coverage, realizing long-time series dynamic interval prediction and point prediction, and integrated output. The construction of its dynamic interval optimization loss function includes:

[0015] S3-1. Prediction interval coverage rate PICP pred K SU =σ(ε1·(y u -y true )) K SL =σ(ε2·(y true -y l )) K S =K SU ·K SL

[0016] Where ε1 and ε2 are smoothing parameters that control the slope of the sigmoid function. SU Use the Sigmoid function to set the upper bound prediction value y u and the true value y true The difference is converted into confidence, indicating the reliability of the upper bound. SL Use the Sigmoid function to calculate the confidence of the lower bound, indicating the lower bound prediction value y l Reliability. S The confidence of the upper and lower bounds is combined by multiplication to obtain the comprehensive confidence K S . This represents the probability that the predicted value falls within both the upper and lower bounds.

[0017] If K SU or K SL If one of them is 0, then K S This means that if the upper bound does not have sufficient confidence, then even if the confidence of the lower bound is high, the interval cannot be considered reliable, and vice versa. This design ensures that the combined confidence K is only valid when both bounds have sufficient confidence. S It enforces that the intervals generated by the model have a certain degree of reliability in both the upper and lower bounds, reflecting a comprehensive assessment of the interval forecast.

[0018] The Sigmoid function enhances the performance of the model in uncertainty processing by converting predicted values ​​into probabilities and provides a smooth gradient to enable it to be transmitted smoothly during the back propagation process.

[0019] Finally get PICP pred , measures the proportion of true values ​​falling within the prediction interval and evaluates the reliability of the model-generated interval.

[0020] S4-2. Average prediction interval width Calculates a weighted mean prediction interval width, ensuring that high confidence intervals contribute more to the width; MPIW pred The smaller the value of , the narrower the prediction interval generated by the model, which means that the model has a better understanding of the distribution of the data; by focusing on the width of high confidence, the model can avoid generating overly broad prediction intervals.

[0021] S4-3, PICP loss L_PICP L PICP =[max(0,(1-α)-PICP pred )] 2 *100 α is the specified confidence level, according to PICPpred The loss is calculated with the target confidence, penalizing cases where the expected coverage is not achieved and encouraging the model to improve the coverage.

[0022] S4-3, Loss beyond range L p For the model prediction value, calculate the loss of whether the actual prediction value exceeds the prediction interval, punish the situation where the actual prediction value exceeds the prediction interval, and enhance the robustness of the model.

[0023] The specific role analysis is as follows: (1) Ensure that the predicted value is within the prediction interval: L p By calculating the predicted value Relative to the upper and lower bounds of the prediction interval y u and l The deviation is penalized for those prediction values ​​that fall outside the prediction interval. hour, The value of is positive, increasing the loss value and generating a penalty; similarly, when the predicted value Below the lower bound y l hour, is positive, increasing the loss value and generating a penalty. The model is encouraged to optimize parameters to reduce these out-of-range predictions so that they are within the prediction interval [y l ,y u ], thereby generating reasonable and reliable interval forecasts. (2) Enforce the model to maintain reasonableness when generating prediction intervals: L p The design of ensures that the model does not generate intervals that are too wide or too narrow. By penalizing those predictions outside the prediction interval, the model will be forced to adjust the width of the interval so that more predictions fall within the reasonable interval. (3) Preventing overly extreme predictions: A prediction interval that is too wide may lead to a low-confidence prediction, while an interval that is too narrow may easily cause the predicted value to fall outside the interval. p The existence of prompts the model to maintain balance and prevent extreme prediction behavior; by controlling the prediction value between the upper and lower bounds, the model avoids over- or under-prediction and enhances its stability in uncertain scenarios.

[0024] S4-5. Mean square error (MSE) MSE is used to evaluate the average difference between the model prediction and the true value, representing the prediction accuracy of the model. In addition to the optimization of interval width (MPIW) and coverage probability (PICP), MSE plays a balancing role to ensure that the model does not sacrifice the accuracy of prediction in order to generate a narrower interval. Deviation from the true value y i , the MSE term will increase the loss, thereby prompting the model to optimize its predicted values ​​to make them closer to the true values.

[0025] To summarize, the total loss of the model is calculated as follows: loss=(1-λ1)(1-λ2)·MPIW pred +λ1(1-λ2)·L PICP +λ2·MSE+ξ·L p λ1, λ2, and ξ play the role of adjusting the weights of different loss terms to balance the model's multiple objectives among prediction accuracy, prediction interval width, coverage probability, and whether the predicted value exceeds the interval range. λ1 controls the prediction interval coverage rate L PICP When λ1 is large, the model tends to optimize coverage and ensure that the true value falls within the prediction interval. When λ1 is small, the model pays more attention to generating a narrow prediction interval. λ2 controls the prediction accuracy of the point. When λ2 is large, the model pays more attention to the accuracy of the predicted value and strives to minimize the difference between it and the true value. When λ2 is small, the model tends to optimize the rationality and width of the interval rather than simply pursuing prediction accuracy. p It is used to penalize the predicted value falling within the interval [y l ,y u ]. A larger ξ will force the model to generate reasonable intervals, ensuring that the predicted value lies between the upper and lower bounds.

[0026] S4. Construct evaluation indicators, divide the training set, validation set and test set, save the model with the best performance on the validation set, apply the saved model to the test set, and calculate the interval coverage and average interval width respectively.

[0027] S4-1, Interval coverage (PICP): Where N is the number of prediction points; a i It is a Boolean parameter, which indicates whether the actual value falls within the predicted interval; i is the predicted target value; L i and U iThe larger the PICP value is, the greater the probability that the interval covers the actual PV output power value.

[0028] S4-2, Mean Interval Width Index (MPIW):

[0029] S5, adaptive adjustment, the model automatically matches the model of the corresponding weather type according to the weather forecast data. According to the different weather type data, the loss function module is loaded into the Transformer model architecture, the hyperparameters of the model are defined, the parameters of the model are updated through back propagation, and the model with the lowest MPIW value under the set confidence level with coverage on the validation set is saved. Weather types can be divided into sunny, cloudy, rainy, etc.

[0030] S6, model reasoning, integrated output of point prediction and interval prediction. The saved model is reasoned and the long time series results are output. The reasoning result is output as the power generation prediction result for the next 24 hours, including the upper limit, lower limit and predicted value. The update interval can be 15 minutes.

Claims

1. A photovoltaic power generation interval prediction method integrating multiple objectives, including model construction, characterized in that: The model construction includes the construction of a dynamic interval optimization loss function. The construction of the dynamic interval optimization loss function includes adopting the Transformer model architecture as the basic model architecture. On the basis of traditional regression tasks, an interval prediction loss function module is designed, and a multi-objective loss function integrating point prediction and interval prediction and a penalty function are proposed. The optimization goal is to minimize the average prediction interval width while maximizing the interval coverage, so as to realize long-time series dynamic interval prediction and point prediction.

2. The photovoltaic power generation interval prediction method integrating multiple objectives according to claim 1 is characterized in that: It also includes feature construction, which includes using a sliding window to construct statistical features of active power and lag features to form a time series data set.

Images