Photovoltaic output prediction method based on multi-feature and multi-model fusion

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a photovoltaic output prediction method based on multi-feature and multi-model fusion, and the method comprises the following steps: S1, collecting the historical output data and related meteorological data of a distributed photovoltaic power station to form a data set, constructing a plurality of single-feature sequences based on the data set, s2, inputting a plurality of combined sequences formed by combining the fixed period feature sequence, the dynamic feature sequence, the local fluctuation feature sequence, the associated meteorological auxiliary features and at least two single feature sequences into an XGBoost model for training and learning to obtain a plurality of single models; s3, performing enhanced fusion on the single models to obtain multiple models, and performing iterative training and optimization on the multiple single models or multiple models to obtain a robust comprehensive integration model; and S4, carrying out probabilistic correction on a prediction result of the comprehensive integration model to obtain a final medium-and-long-term distributed photovoltaic prediction sequence. The method has the advantage that the prediction capability of the provided model framework has advancement and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic output prediction method based on the fusion of multiple features and multiple models. Background Art

[0002] Photovoltaic power generation is a green and environmentally friendly new energy power generation method. The vigorous development of the photovoltaic industry is of great significance to promoting the high-quality development of China's economy, building a low-carbon and efficient energy system, and ensuring the realization of the "dual carbon" goal. my country's photovoltaic industry has developed very rapidly, and the newly installed capacity has also increased year by year. In 2023, the domestic photovoltaic installed capacity will be 216.88GW, a year-on-year increase of 148.1%, of which the new installed capacity of centralized photovoltaics will exceed 110GW, and the new installed capacity of distributed photovoltaics will also be nearly 100GW. From the perspective of time and space, the new installed capacity in China in 2023 is equivalent to the sum of the new installed capacity in China from 2019 to 2022, and basically exceeds the sum of the increments in other countries and regions in the world in that year. Distributed photovoltaic power stations can be dispersed in various regions, can directly supply electricity demand locally, and reduce energy consumption and losses in the transmission process. If the medium- and long-term photovoltaic output can be predicted in a timely and accurate manner, it will have important value for applications such as light resource assessment, power grid operation, and planning of new photovoltaic power stations. Photovoltaic power stations are a typical intermittent power source; and affected by meteorological and environmental conditions, they have strong diurnal and seasonal periodicity, many influencing factors, complex mechanisms, strong volatility and randomness, and fast frequency of change. These factors bring great difficulties to the accurate prediction of medium- and long-term distributed photovoltaic output. In some existing studies, the modeling methods based on mechanism models, the modeling methods based on probability statistics, the modeling methods based on artificial intelligence algorithms, and other direct or indirect prediction methods are mainly used to carry out research.

[0003] The mechanism modeling method mainly considers physical equations such as solar radiation transfer equation and photovoltaic component operation equation for prediction research. It does not require a large amount of historical data, but requires detailed geographic information, component parameters, and meteorological and solar radiation data of photovoltaic power stations. It is suitable for newly built photovoltaic power stations. The modeling process is complicated and it is difficult to simulate some extreme abnormal weather conditions and the slow changes of environmental and photovoltaic component parameters over time. The model has poor anti-interference ability and is not robust. The main prediction models include the Clear Sky Model (CSM), Solis model, ESRA model, Bird and Hulstrom model, Ineichen model, etc. Under clear sky conditions, photovoltaic power fluctuations are small, which can maximize the reflection of the power generation effect of irradiance. Therefore, the clear sky prediction model is often used as the basic calibration model. Cloud movement is the main reason for the fluctuation of ground irradiance and photovoltaic power. The prediction method based on cloud map has become one of the important technical directions for refined photovoltaic prediction. All cloud maps need to be processed by image processing to identify clouds, and block matching technology, cross-correlation algorithm and other methods are used to predict cloud movement, form cloud index map, and predict cloud occlusion. The analysis data of satellite cloud images is combined with ground telemetry data, and used as the input of the prediction model for prediction. The ground-based cloud image data, satellite cloud image data, and NWP data are combined for prediction. The prediction method based on probability statistics uses the geographical information of the power station, numerical weather forecasts, meteorological station observation information, and long-term power station operation information for a large amount of learning to establish a prediction model based on probability statistical analysis and multi-mode collection; it mainly includes gray theory, time series method, and probability prediction method, etc., and uses parameter estimation and curve fitting to obtain the correspondence between historical meteorological data and power generation, so as to realize the prediction of photovoltaic power station power generation. In recent years, with the gradual maturity of artificial intelligence technology and big data analysis technology, traditional modeling methods have been gradually replaced by advanced technologies, especially artificial intelligence modeling methods represented by machine learning and deep learning technology, such as ensemble learning model (Ensemble Model), recurrent neural network model (Recurrent Neural Network, RNN), long short-term memory network model (Long Short-Term Memory, LSTM), Transformer network model, particle swarm algorithm, genetic algorithm, etc., all of which have achieved good results.

[0004] Current research mostly considers the impact of meteorological data on power generation, while ignoring the periodic variation characteristics of photovoltaic power generation data at different time scales. There are few comprehensive modeling and analysis studies on the output characteristics of photovoltaic power stations in different regions and the dynamic analysis of historical time series characteristics. At the same time, due to the limitations of algorithm structure and strategy, it is difficult to effectively integrate multiple models for accurate prediction on a long time scale. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a photovoltaic output prediction method based on the fusion of multiple features and multiple models.

[0006] The object of the present invention is achieved by the following technical solution: A photovoltaic output prediction method based on multiple features and multiple models fusion, comprising the following steps:

[0007] Step S1: Collect historical output data of distributed photovoltaic power stations and related meteorological data to form a data set, and construct multiple single feature sequences based on the data set, including fixed period feature sequences of different time scales, dynamic feature sequences, local fluctuation feature sequences and associated meteorological auxiliary features;

[0008] Step S2: inputting multiple combined sequences formed by combining fixed period feature sequences, dynamic feature sequences, local fluctuation feature sequences, associated meteorological auxiliary features, and at least two single feature sequences into the XGBoost model for training and learning to obtain multiple single models;

[0009] Step S3: Enhance and fuse single models with different features and parameters to obtain multiple models, iteratively train and optimize multiple single models or multiple models to obtain a robust comprehensive integrated model;

[0010] Step S4: Probabilistically correct the prediction results of the comprehensive integrated model to obtain the final medium- and long-term distributed photovoltaic prediction sequence.

[0011] Specifically, the fixed period feature sequence is:

[0012] C={v(dn,h)n=1,2,…,m; h∈[1,96]}

[0013] Where: v(dn,h) represents the output value at the same time point with a delay of n days, and n is a positive integer.

[0014] Specifically, the method for constructing the dynamic feature sequence is:

[0015] In the training phase, the sliding window size is set, the current pattern string is determined, the used historical sequence is split into multiple pattern sequences and traversed in sequence, and the sequence similarity values ​​except the last position data point are calculated based on the similarity algorithm, the largest K sequences are selected, and the last position data point is taken to form a dynamic feature sequence. The label is designed to correspond to the time point feature of the sequence, and finally combined into a pattern library Pattern_Tag with different labels for the corresponding station;

[0016] In the prediction stage, according to each prediction time point i, a pattern string pattern_i to be matched is combined, and the most relevant feature sequence D_i is selected from Pattern_Tag based on similarity calculation.

[0017] Specifically, the local volatility characteristic sequence is: W_t = {FA, FR, RR, σ, A, PP, PM};

[0018] Among them, FA is the fluctuation amplitude; FR is the volatility; RR is the climbing rate; σ is the fluctuation variance; A is the maximum fluctuation amplitude;

[0019] γ i =v i+1 -v i (1)

[0020]

[0021] A=P max -P min (8)

[0022] Among them, γ i represents the photovoltaic output value at adjacent moments, n is the statistical window size; t j , t i represents the j, i time point, v j , v i represents the photovoltaic output value at time point j, i, ρ m represents the climbing rate of a certain climbing; R(i,j) represents the number of times the climbing condition is met within the statistical time window (i,j), N represents the number of climbing conditions in a period of time, and RR represents the average climbing rate in a period of time; P(i) represents the actual photovoltaic output value at the i-th moment, Represents the average value of photovoltaic output during the statistical period, P max , P min They represent the maximum and minimum values ​​of photovoltaic output respectively; the fluctuation variance σ reflects the degree to which the photovoltaic output deviates from its average value at any time within a period of time. The larger the fluctuation variance, the more violent the fluctuation; the maximum fluctuation amplitude A reflects the maximum range of photovoltaic output within a period of time. The larger A is, the greater the fluctuation is.

[0023] Specifically, the associated meteorological auxiliary characteristics include downward solar shortwave radiation reaching the ground, all-sky direct solar radiation received by the ground, and direct solar radiation meteorological characteristics.

[0024] Specifically, in step S3, the probabilistic multi-model fusion method based on the optimal scoring matrix is ​​used to fuse the prediction results of multiple models. First, the number of times the accuracy of each model is optimal in each stage is counted as the model selection score dependency value of the corresponding stage. All models in all time periods form an optimal scoring matrix. Then, in the training stage, the optimal scoring matrix is ​​constructed according to the backtest results. In the prediction stage, the model with the maximum probability is selected from the matrix according to the time parameter.

[0025] The present invention has the following advantages:

[0026] The present invention takes into account the volatility of photovoltaic output and the differences in data distribution in different regions, designs multiple feature combinations, and increases feature expression and adaptability. Based on the Boost ensemble learning model, single models of different time scales are trained, and the optimal scoring strategy is used for multi-model fusion in the rolling prediction process. Finally, a probabilistic correction is performed for each time period. Experiments show that the effectiveness of M2E-DPV under a variety of feature combinations and multi-model fusion strategies, the average MAPE, R2, and ACC indicators are 0.15, 0.96, and 0.91, respectively, which are greatly improved compared with other methods, indicating that the prediction ability of the model framework proposed in the present invention is advanced and robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the flow chart of the output prediction method of the present invention;

[0028] Figure 2 A schematic diagram of the fixed period feature sequence of the present invention

[0029] Figure 3 Schematic diagram of the DTW algorithm of the present invention;

[0030] Figure 4 The modeling process and framework diagram of the present invention;

[0031] Figure 5 It is a trend chart of the medium- and long-term distributed photovoltaic output forecast capability;

[0032] Figure 6 This is a comparison curve of medium- and long-term distributed photovoltaic output. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0034] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0035] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0036] The present invention is further described below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following description. Figures 1 to 6 As shown, a photovoltaic output prediction method based on the fusion of multiple features and multiple models includes the following steps: Step S1: collecting historical output data and related meteorological data of distributed photovoltaic power stations to form a data set, and constructing multiple single feature sequences based on the data set, including fixed period feature sequences of different time scales, dynamic feature sequences, local fluctuation feature sequences and associated meteorological auxiliary features;

[0037] The fixed period feature sequence means that multiple historical values ​​are continuously selected as input features from the historical sequence with a certain step length as the period; according to the analysis of photovoltaic output characteristics and autocorrelation sequence, the correlation coefficients of the same time in adjacent days are all above 0.8. Therefore, the day is selected as the step length (96 time points), and m steps (m=12 in the present invention) are selected forward to construct a fixed period feature sequence, such as Figure 2 As shown in the figure, P(d,h) represents the predicted photovoltaic output value at the hth time point on the dth day, v(dn,h) represents the output value at the same time point with a delay of n days, where n is a positive integer, and the periodic feature sequence C = {v(dn,h)n = 1,2,…,m; h∈[1,96]} is obtained; the fixed periodic feature expresses the data change trend under a specific time scale;

[0038] Affected by the weather and geographical location, photovoltaic output has certain volatility and randomness. The present invention proposes an adaptive dynamic feature construction method to cope with the impact of photovoltaic output changes. The dynamic feature is based on the current pattern string pattern_t (defined as a sequence of continuous data points with a certain time span in the historical sequence close to the current time t). Through similarity calculation, a historical output sequence (such as top K) with high trend and numerical change consistency with pattern_t is matched from the historical pattern library, and the value of the last position is selected to form an input feature sequence D_t = {d_i, i = 1, 2, ..., K}, where K is the feature length. The dynamic feature selection is based on an assumption that at the same time scale, the sequence matched based on similarity has a consistent change trend and numerical law of data points in most time periods, so the change of a single data point at a local position should also be consistent. The construction process of the dynamic feature is as follows:

[0039] In the training phase, the sliding window size (such as 96) is set, the current pattern string is determined (a certain time scale is set, such as an interval of 1 time point or 96 time points, etc.), the used historical sequence is split into multiple pattern sequences and traversed in sequence, and the sequence similarity values ​​except the last position data point are calculated according to the similarity algorithm, and the largest K (indicating the dynamic feature length) sequences are selected. The data point at the last position is taken to form a dynamic feature sequence. The label is designed to correspond to the time point feature of the sequence (such as month-day-time point), which expresses the consistency of photovoltaic output data at different times, and finally combined into a pattern library Pattern_Tag with different labels for the corresponding site;

[0040] In the prediction stage, according to each prediction time point i, the pattern to be matched pattern_i is combined according to the same time scale, and the most relevant feature sequence D_i is selected from Pattern_Tag based on the similarity calculation, that is, D_i = Similarity(Pattern_Tag, pattern_i). Similarity is a certain type of similarity calculation method. In actual scenarios, Euclidean distance, cosine distance, dynamic time warping (Dynamic Time Warping, DTW) algorithm, etc. are generally selected. The present invention mainly uses the DTW algorithm to calculate pattern similarity.

[0041] The basic idea of ​​the DTW algorithm is to measure the similarity of two time series by calculating the minimum distance between them, and to construct the correspondence between the elements of two sequences (A / B) of different lengths according to the principle of the shortest distance. The construction process is mainly divided into two parts: constructing the cumulative distance matrix and finding the shortest path. When constructing the correspondence between the elements of two sequences, the sequences need to be extended or compressed, such as Figure 3 As shown;

[0042] Local volatility characteristics refer to the statistical characteristics of the peak and valley values, fluctuation variance, fluctuation amplitude, climbing times, etc. of photovoltaic output in a period of time before the forecast target, which describe the extreme value and stability of output changes. The definition and calculation of the main volatility characteristics are as follows:

[0043] Fluctuation Amplitude FA and Volatility Rate FR

[0044] γ i =v i+1 -v i (1)

[0045]

[0046] Among them, γ i It represents the photovoltaic output value at adjacent moments, and n is the size of the statistical window. The fluctuation amplitude FA and the fluctuation rate FR reflect the fluctuation degree of photovoltaic output at adjacent moments, which are usually calculated by the average value within a time window. The smaller the value, the smaller the fluctuation and the better the stability.

[0047] Climbing rate RR

[0048]

[0049] Among them, t j , t i represents the j, i time point, v j , v i represents the photovoltaic output value at time point j, i, ρ m It represents the climbing rate of a certain climbing; R(i,j) represents the number of times the climbing condition is met within the statistical time window (i,j), N represents the number of climbing conditions in a period of time, and RR represents the average climbing rate in a period of time. The climbing rate fully describes the fluctuation process of the photovoltaic output time series.

[0050] Volatility variance σ and maximum volatility A

[0051]

[0052] A=P max -P min (8)

[0053] in, represents the average value of photovoltaic output during the statistical period, P(i) represents the actual photovoltaic output value at the i-th moment, and P max , P minRespectively represent the maximum and minimum values ​​of photovoltaic output; the fluctuation variance σ reflects the degree to which photovoltaic output deviates from its average value at any time within a period of time. The larger the fluctuation variance, the more violent the fluctuation; the maximum fluctuation amplitude A reflects the maximum range of photovoltaic output within a period of time. The larger A is, the greater the fluctuation is. In addition, the peak-to-valley average range PP, the peak-to-valley mean PM, etc. are also used as volatility feature sequences, and the photovoltaic output volatility feature sequence W_t = {FA, FR, RP, σ, A, PP, PM} at time t is obtained; associated meteorological auxiliary features, meteorological factors have the most direct impact on photovoltaic output, especially shortwave, solar irradiance and other meteorological elements; through correlation coefficient analysis, the downward solar shortwave radiation reaching the surface (Surface Short-wave Radiation Downwards, SSRD), the total sky direct solar radiation received by the ground (Total Sky Direct Solar Radiation at Surface, FDIR), and direct solar radiation (Direct Solar Radiation, DSRP) are selected.

[0054] Table 1 Main meteorological factors affecting photovoltaic output

[0055]

[0056] In addition, One-Hot coding feature conversion is performed on season, month, and time point, where the coding length of season is 4, the coding length of month is 12, and the coding length of time point is 96.

[0057] Step S2: inputting multiple combined sequences formed by combining fixed period feature sequences, dynamic feature sequences, local fluctuation feature sequences, associated meteorological auxiliary features, and at least two single feature sequences into the XGBoost model for training and learning to obtain multiple single models;

[0058] XGBoost Model:

[0059] The present invention adopts the eXtreme Gradient Boosting (XGBoost) algorithm based on the Boost integrated architecture as the medium- and long-term distributed photovoltaic output prediction learning framework, the Boosting integrated technology, the basic idea is to combine multiple base (weak) learners for serial training, the latter base learner corrects the error of the previous base learner, and finally all base learners are integrated into a strong learner. The model function based on the Boost integrated learning architecture is expressed as:

[0060]

[0061] where x i ,y i are the independent variable and dependent variable of the sample set, respectively, and f k is a base learner (the mapping function between independent variables and dependent variables), K is the number of base learners (decision trees are usually used as base learners), and the basic idea is to use the distributed forward additive model to iteratively improve and generate base learners by optimizing the structured loss function.

[0062] The principle of XGBoost model is as follows:

[0063] First, construct the objective function and introduce the regularization term into the loss function to obtain the basic objective function:

[0064]

[0065] In the formula, the first term L(y i ,y) represents the loss function of the model, and the true value of each sample y i The loss between the predicted value y and the second term Ω(f k ) represents the complexity of the tree, such as the number of tree layers, the number of leaf nodes, etc. Since the algorithm uses superposition training, that is, the prediction value of k models is equal to the prediction value of the previous k models plus the prediction value of the kth model currently being trained, so:

[0066]

[0067] Introducing the objective function, we get:

[0068]

[0069] For the kth tree being trained, the first k-1 trees are all constants, so the objective function becomes:

[0070]

[0071] in, is the predicted value of the first k-1 tree models, which is also a constant. The Taylor series approximates the objective function and is defined as:

[0072]

[0073] Because g i and h i They are all related to the first k-1 learners, so they are all constants. is a constant and has nothing to do with optimizing the kth tree, so omitting the constant term yields:

[0074]

[0075] Then the tree structure is introduced into the objective function to optimize the unknown quantity f k (x i ) and Ω(f k ), it needs to be parameterized to obtain the parameterized objective function:

[0076]

[0077] Among them, T represents the number of leaf nodes, represents the sum of the values ​​of each leaf node, α and λ are hyperparameters that control the degree of penalty. After introducing the parametric expression of the tree structure, a greedy algorithm is used to construct K optimal trees (to minimize the objective function), and finally all the trees are weighted to form our final model:

[0078]

[0079] Step S3: Enhance and fuse single models with different features and parameters to obtain a multi-model, iteratively train and optimize multiple single models or multi-models to obtain a robust comprehensive integrated model.

[0080] Different features and parameters refer to the use of different feature combinations. In the process of using different features and training models, certain combinations are adopted. Some types of features have special treatments. For example, for fixed periods, different time scales will be considered, such as seasonal trends, short-term trends, etc.

[0081] The rolling forecast method is used for medium- and long-term forecast output. Since the feature input in the subsequent forecast process may use the output of the previous forecast process, it is easy to introduce additional uncertainty and cause error accumulation. Train multiple single models with different time scales, fuse the prediction results of multiple models in each step of the forecast, and reduce the accumulation of errors through model adaptation.

[0082] The present invention proposes a probabilistic multi-model fusion method based on the optimal scoring matrix. The basic idea is to calculate the degree of dependence on models of different scales at different time points, and to count the number of times each model has the best accuracy at each stage (or time label) as the model selection scoring dependency value for the corresponding stage. All models in all time periods form the optimal scoring matrix. In the training stage, the optimal scoring matrix is ​​constructed based on the backtest results, and in the prediction stage, the model with the maximum probability is selected from the matrix based on the time parameter. The basic process is as follows:

[0083] Training phase

[0084] Input: model set S, backtest data set RD

[0085] Output: optimal rating matrix M

[0086] Step 1: Initialize the rating matrix M. According to the monthly dimension, there is one matrix for each month. The rows represent the model mid and the columns represent the time point. Initially, all the matrix are set to 0.

[0087] Step 2: For the trained model set S, backtest the training set RD respectively, and calculate the residuals between the predicted values ​​and the true values ​​of all models at all time points;

[0088] Step 3: According to the size of the residual value, select the model at the time point corresponding to the minimum value, and update the number of times of the corresponding position in M, that is, M[mid_i, timepoint_i] + = 1;

[0089] Step 4: After all models are backtested and updated at all time points, the number of selections in the normalized matrix M is used as the score, and the optimal score matrix is ​​obtained.

[0090] Prediction Phase

[0091] Input: optimal rating matrix M, test data set PD

[0092] Output: Prediction sequence

[0093] Step 1: Match the position of the current prediction time point i in M ​​[*,i], that is, the column where the prediction time point i is located;

[0094] Step 2: Calculate the probability of each model selection in the column where the prediction time point is located;

[0095] Step 3: According to the probability maximization principle, select the position with the largest probability value, obtain the model of the corresponding position as the current prediction model, and output the predicted value.

[0096] Among them, the calculation of probability maximization adopts the softmax function, that is,

[0097]

[0098] m i It is the element value of the model corresponding to the score vector at the i-th time point. The function of the softmax function is to convert each element in the input vector into its form as a probability value.

[0099] Step S4: The prediction results of the comprehensive integrated model are corrected in a probabilistic manner by time period to obtain the final medium- and long-term distributed photovoltaic prediction sequence.

[0100] The present invention takes into account the volatility of photovoltaic output and the differences in data distribution in different regions, and designs various types of features, such as fixed period features, dynamic features, local volatility features, and related meteorological auxiliary features. A variety of feature sequences are comprehensively modeled to increase the expressiveness and adaptability of features. Different combinations have a certain impact on the results. Considering the large medium and long-term time span, in order to avoid the influence of error accumulation during the iteration process, a probabilistic multi-model fusion method based on the optimal scoring matrix is ​​adopted. According to the statistical characteristics of photovoltaic output of different sites, the fusion results are probabilistically corrected in different time periods. Photovoltaic power stations in two regions are modeled separately, and the MAPE indicators are improved by 6.2% and 10.5% respectively compared with the benchmark strategy. Finally, the proposed prediction framework and strategy are compared from the dimensions of different regions, different feature combinations, and different methods, which shows that the prediction ability of the model framework proposed by the present invention is advanced and robust.

[0101] The present invention verifies the proposed model and strategy from different dimensions by designing multiple groups of experiments. First, the historical data of distributed photovoltaic and related data are collected and sorted, and the training set and test set are divided, and evaluation indicators are designed; then, a comparison is made from the dimensions of multi-feature combination and multi-region; then, a variety of currently popular classic prediction methods are compared on the same data set, such as SVM, LSTM, ET, Transformer, LightGBM, etc., and finally the results are further analyzed and discussed.

[0102] Datasets,

[0103] The present invention uses historical output data of three distributed photovoltaic power stations in a certain area of ​​Sichuan Province and three distributed photovoltaic power stations in a certain area of ​​Shanxi Province for experiments. The time range of historical data is from the beginning of 2022 to the beginning of 2024, with a total of 2 full years of data, and the time resolution is 15 minutes. The installed capacity of several photovoltaic power stations is equivalent; the design experiment selects the photovoltaic output of four typical months (March, June, September, and December) in 2023 as the prediction target, that is, the first day of the month at 0 time is used as the starting time, and the photovoltaic output at each time point in the next 30 days is predicted. The forecast is made continuously for 30 days, and the average forecast ability of each day is counted. A total of 2880 time points are forecasted in a single batch. The training sets are divided according to all historical time period data before their respective starting times, such as the first group of training sets from 2022-01-01 00:15 to 2023-03-01 00:00; other groups are similar.

[0104] In addition, considering the ECMWF (European Centre for Medium-Range Weather Forecasts) meteorological source forecast data for the same period as the auxiliary features, three types of features with greater correlation at the grid locations (about 9km*9km) near the corresponding photovoltaic power station are selected, namely, downward solar shortwave radiation (SSRD) reaching the surface, full-sky direct solar radiation (DSRP) received on the ground, and direct solar radiation (FDIR).

[0105] Evaluation Metrics

[0106] Model evaluation is one of the important means of testing performance. It measures the predictive ability and stability of the model from multiple perspectives. The main evaluation indicators used in this paper are MAPE (mean absolute percentage error), MAE (mean absolute error), R2 (coefficient of determination), and ACC_new (accuracy), which are defined as follows:

[0107]

[0108] Where N is the number of evaluation samples, y i is the actual value, is the predicted value, is the mean of the actual values, Cap is the installed capacity of the PV power station, and MAPE is the average percentage of the relative error between the predicted value and the actual value, with a value range of [0, +∞]. The smaller the value, the more accurate the prediction model is; R 2 It expresses the goodness of fit calculation of a model. The closer to 1, the better the fit effect. ACC is the accuracy rate. The closer the value is to 1, the better.

[0109] Analysis of multiple feature combination results

[0110] The performance of different features and their combinations are compared. For the convenience of display, the fixed period feature is represented as T1, the dynamic feature is represented as T2, the local volatility feature is represented as T3, the meteorological feature is represented as T4, and the time coding feature is represented as T5; a variety of feature combination schemes are designed, such as T1+T4, which represents the combination of fixed period features and meteorological features, and the others are similar. SX region and SC region represent Shanxi region and Sichuan region respectively. The results of multiple photovoltaic power stations in the same region in the four typical months of the experiment (March, June, September, and December) are averaged. The results of the present invention are carried out according to the basic rolling prediction method. The specific results are shown in the following table:

[0111] Table 2 Comparison of different feature combinations in the same model

[0112]

[0113] By comparing the results of multiple features and their combinations, in the case of single features, T1 and T2 have comparable effects, and the average MAPE values ​​in the SX and SC regions are both 0.19, which is better than the single T3 feature result; in terms of accuracy index comparison, the two regions also show consistent results. In the case of combined features, the five features are combined, and the index performance is the best under the set experimental conditions, among which the average MAPE in the SX region is 0.16, R2 is 0.94, and ACC is 0.85; the average MAPE in the SC region is 0.19, R2 is 0.92, and ACC is 0.84, which are better than other combinations; in particular, the combination of T2 and T4 is more ideal. From the average results of distributed photovoltaic power stations in different regions, the overall performance of the SX region is slightly better than that of the SC region. It shows that the designed combined features are effective in photovoltaic output prediction. By comparing the average conditions of multiple groups of different feature combinations and photovoltaic power stations in different regions, it is shown that the stability of the combined features is also better than that of single features.

[0114] Comparison of results from different models

[0115] The model framework results proposed in this paper are compared with some popular models in current research. Models such as LSTM, ET, SVM, Transformer, RF, and LightGBM are selected to further illustrate the advantages of the model framework and strategy proposed in this paper. It should be noted that in the experimental verification of different models, the combined features (T1+T2+T3+T4+T5) designed in Section 4.3 are fixedly used as the basic feature input of all models, and the optimal scoring strategy is used for iteration. The comparison of the results of various models is shown in the following table:

[0116] Table 3 Comparison of different models for the same set of features

[0117]

[0118] Among them, M2E-DPV * It is the abbreviation of the main model framework and strategy proposed in this paper. From the indicators of the two regions, M2E-DPV * Compared with other comparative models, several comparative indicators of the method have significantly improved performance, and the indicator results in different regions are more stable, which further illustrates the advantages of the model framework strategy proposed in this invention.

[0119] Finally, according to the above experimental results, the forecasting capabilities of the two regions are statistically analyzed and the forecast trends are displayed, such as Figure 5 ,in, Figure 5 The middle left and right parts are the forecast capability statistics of the three indicators for SX and SC regions for 30 consecutive days. Figure 6In the middle, a comparison chart of the predicted value and actual value is extracted with a certain time point as the starting time and predicted for the next 30 days (30*96 data points).

[0120] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above-mentioned technical content without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any changes, modifications, equivalent changes and modifications made to the above embodiments based on the technology of the present invention without departing from the content of the technical solution of the present invention belong to the protection scope of the present technical solution.

Claims

1. A photovoltaic output prediction method based on multiple features and multiple models fusion, characterized by: The following steps are involved: Step S1: Collect historical output data of distributed photovoltaic power stations and related meteorological data to form a data set, and construct multiple single feature sequences based on the data set, including fixed period feature sequences of different time scales, dynamic feature sequences, local fluctuation feature sequences and associated meteorological auxiliary features; Step S2: inputting multiple combined sequences formed by combining fixed period feature sequences, dynamic feature sequences, local fluctuation feature sequences, associated meteorological auxiliary features, and at least two single feature sequences into the XGBoost model for training and learning to obtain multiple single models; Step S3: Enhance and fuse single models with different features and parameters to obtain multiple models, iteratively train and optimize multiple single models or multiple models to obtain a robust comprehensive integrated model; Step S4: Probabilistically correct the prediction results of the comprehensive integrated model to obtain the final medium- and long-term distributed photovoltaic prediction sequence.

2. The photovoltaic output prediction method based on multiple features and multiple models fusion according to claim 1 is characterized by: The fixed period characteristic sequence is: C={v(dn,h)n=1,2,…,m; h∈[1,96]} Where: v(dn,h) represents the output value at the same time point with a delay of n days, and n is a positive integer.

3. The photovoltaic output prediction method based on multiple feature and multi-model fusion according to claim 1 is characterized by: The method for constructing the dynamic feature sequence is: In the training phase, the sliding window size is set, the current pattern string is determined, the used historical sequence is split into multiple pattern sequences and traversed in sequence, and the sequence similarity values ​​except the last position data point are calculated based on the similarity algorithm, the largest K sequences are selected, and the last position data point is taken to form a dynamic feature sequence. The label is designed to correspond to the time point feature of the sequence, and finally combined into a pattern library Pattern_Tag with different labels for the corresponding station; In the prediction stage, according to each prediction time point i, a pattern string pattern_i to be matched is combined, and the most relevant feature sequence D_i is selected from Pattern_Tag based on similarity calculation.

4. The photovoltaic output prediction method based on multiple feature and multi-model fusion according to claim 1 is characterized in that: The local volatility characteristic sequence is: W_t = {FA, FR, RR, σ, A, PP, PM}; Among them, FA is the fluctuation amplitude; FR is the volatility; RR is the climbing rate; σ is the fluctuation variance; A is the maximum fluctuation amplitude; γ i =v i+1 -v i (1) A=P max -P min (8) Among them, γ i represents the photovoltaic output value at adjacent moments, n is the statistical window size; t j , t i represents the j, i time point, v j , v i represents the photovoltaic output value at time point j, i, ρ m represents the climbing rate of a certain climbing; R(i,j) represents the number of times the climbing condition is met within the statistical time window (i,j), N represents the number of climbing conditions in a period of time, and RR represents the average climbing rate in a period of time; P(i) represents the actual photovoltaic output value at the i-th moment, Represents the average value of photovoltaic output during the statistical period, P max , P min They represent the maximum and minimum values ​​of photovoltaic output respectively; the fluctuation variance σ reflects the degree to which the photovoltaic output deviates from its average value at any time within a period of time. The larger the fluctuation variance, the more violent the fluctuation; the maximum fluctuation amplitude A reflects the maximum range of photovoltaic output within a period of time. The larger A is, the greater the fluctuation is.

5. The photovoltaic output prediction method based on multiple feature and multi-model fusion according to claim 1 is characterized in that: The associated meteorological auxiliary characteristics include downward solar shortwave radiation reaching the ground surface, all-sky direct solar radiation received by the ground, and direct solar radiation meteorological characteristics.

6. The photovoltaic output prediction method based on multiple feature and multi-model fusion according to claim 1 is characterized in that: In step S3, the probabilistic multi-model fusion method based on the optimal scoring matrix is ​​used to fuse the multi-model prediction results. The method first counts the number of times each single model has the best accuracy in each stage as the model selection score dependency value in the corresponding stage, then constructs the optimal scoring matrix according to the backtest results in the training stage, and selects the model with the maximum probability from the matrix according to the time parameter in the prediction stage.

Images