A photovoltaic cluster power prediction method based on similar station correlation feature mining

Abstract

The application discloses a photovoltaic cluster power prediction method based on similar station related feature mining, and relates to the technical field of photovoltaic power prediction.The method comprises the following steps: dividing a to-be-predicted day into a sunny day prediction day and a non-sunny day prediction day; for the sunny day prediction day, directly using a pre-trained time convolution network model for prediction; for the non-sunny day prediction day, constructing a historical theoretical power difference value of a target station, simultaneously screening similar stations according to power generation and weather characteristics, determining historical similar weather characteristics, inputting into a time convolution neural network, predicting a power difference value, and calculating a non-sunny day power prediction result; and performing power prediction on all photovoltaic stations in a photovoltaic cluster to obtain a predicted power of the photovoltaic cluster. By analyzing weather characteristics of adjacent stations and calculating a historical theoretical power difference value of a target station, a time convolution network model is applied to capture long-term dependence in a sequence, and the prediction accuracy and prediction reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power prediction technology, and in particular to a photovoltaic cluster power prediction method based on the mining of relevant features of similar power plants. Background Technology

[0002] With the continuous increase in installed capacity of new energy sources, the transformation of the world's energy structure is accelerating. Due to the randomness, intermittency, and volatility of photovoltaic power generation, as its scale continues to expand, its proportion in the energy structure is increasing, and its impact on the safe and stable operation of the power system will become increasingly significant. In order to ensure regional power balance, large-scale photovoltaic cluster forecasting is imperative.

[0003] Existing short-term photovoltaic power forecasts mostly use irradiance and temperature as the main features for modeling. However, the usability of irradiance forecasts under severe weather conditions is poor, the mapping ability of the forecasting models is not accurate enough, and the selection of modeling features usually focuses on selecting the time series data with the highest power correlation related to weather, lacking consideration of parameters such as cloud cover, precipitation, and humidity, thus wasting valuable meteorological station resources. The forecasting of extreme fluctuating weather is usually based on strategies such as clustering of similar days, model performance optimization, feature extraction, or error correction. These strategies often struggle to balance forecasting accuracy and modeling cost and have a high degree of randomness. The impact of multi-source site information on photovoltaic power forecasting is also neglected. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic cluster power prediction method based on the mining of relevant features of similar power plants, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants includes:

[0007] The historical photovoltaic power and numerical weather forecast data of the target site are obtained, and the forecast days are divided into clear weather forecast days and non-clear weather forecast days.

[0008] For the sunny day prediction, the pre-trained temporal convolutional network model is used directly to make the prediction and obtain the sunny day power prediction result.

[0009] For non-sunny forecast days, the historical theoretical power difference of the target power station is detrended. At the same time, similar power stations are selected based on power generation and weather characteristics to determine historical similar meteorological characteristics. The historical theoretical power difference and historical similar meteorological characteristics are input into a temporal convolutional neural network to predict the power difference and calculate the non-sunny power prediction result of the target power station.

[0010] Power prediction is performed on all photovoltaic power stations in the photovoltaic cluster to obtain the predicted power of the photovoltaic cluster.

[0011] Furthermore, the historical theoretical power difference of the target power station is constructed to de-trend, including:

[0012] The real power in the historical data of the target power station is searched according to the preset search range to construct the clear sky photovoltaic curve;

[0013] Based on the clear-sky photovoltaic curve, calculate the historical theoretical power difference of the photovoltaic cluster detrending.

[0014] Furthermore, based on a preset search interval, the actual power output in the historical data of the target power station is searched to construct a clear-sky photovoltaic curve, including:

[0015] The solar photovoltaic curve under clear skies is expressed as follows:

[0016]

[0017] In the formula, Let be the clear sky power value at time point i in the k-th search interval; Let be the actual power value at time point i of the power data on day m in the k-th search interval; m is the number of days in each search interval.

[0018] Furthermore, based on the clear-sky photovoltaic curve, the historical theoretical power difference of the photovoltaic cluster detrending is calculated, including:

[0019] Based on the clear-sky photovoltaic curve, the historical theoretical power difference of the target power station that has not detrended is calculated, and the expression is:

[0020]

[0021] In the formula, The undetrended historical theoretical power difference is the value of time point i on day m in the k-th search interval; Let i be the daily net air power value at time point i in the k-th search interval; The actual power value at time point i of the power data on day m in the k-th search interval;

[0022] Based on the undetrended historical theoretical power difference, the detrended historical theoretical power difference is calculated, expressed as:

[0023]

[0024] In the formula, It represents the detrended historical theoretical power difference of time point i on day m in the k-th search interval.

[0025] Furthermore, based on power generation capacity and weather characteristics, similar power stations are screened to determine a set of similar meteorological characteristics, including:

[0026] Search for a set of similar stations to the target station based on dynamic time-normalized distance;

[0027] By filtering historical weather characteristics of similar stations based on cloud cover and precipitation, a set of similar meteorological characteristics is obtained.

[0028] Furthermore, based on the dynamic time-warped distance, a set of similar stations for the target station is searched, including:

[0029] Construct the distance matrix M between the target station and the reference station, including:

[0030] Given that the target site has m time points, the first time series is: The reference station has n time points, resulting in the second time series. Calculate the Euclidean distance between different time points and generate the distance matrix M, expressed as:

[0031]

[0032]

[0033] In the formula, The distance is Euclidean. The power generation at the i-th time point in the first time series; The power generation at the j-th time point in the second time series;

[0034] Calculate the minimum cumulative distance based on the distance matrix M, and generate the cumulative distance matrix W, expressed as:

[0035]

[0036]

[0037] In the formula, The minimum cumulative distance required to walk from the starting point to position (i,j); The Euclidean distance between the two time points;

[0038] Based on the cumulative distance matrix W, the DTW distance between the target station and the reference station is obtained, expressed as:

[0039]

[0040] In the formula, The DTW distance between the first and second time series is represented by a smaller value, indicating that the invention modes of the target site and the reference site are more similar. The value at the top right corner of the cumulative distance matrix W represents the total cost of optimal alignment from beginning to end;

[0041] The DTW distance between the target station and the reference stations is used to generate DTW vectors with multiple reference stations, expressed as follows:

[0042]

[0043] In the formula, Let be the DTW distance between the i-th target station and the n-th reference station; The time series of the i-th target station; This is the time series of the nth reference station;

[0044] Selecting the k most similar reference stations yields the set of similar stations, expressed as:

[0045]

[0046] In the formula, Let i be the set of similar stations to the i-th target station; Let k be the kth similar station with the smallest DTW distance; k is the number of similar stations.

[0047] Furthermore, based on cloud cover and precipitation, historical weather characteristics of similar stations were filtered to obtain a set of similar meteorological characteristics, including:

[0048] Weather classification was performed on historical data from similar stations based on cloud cover and precipitation to obtain cloud cover time series and precipitation time series;

[0049] A day with cloud cover less than 30% of the sky is considered sunny; a day with cloud cover between 30% and 70% is considered cloudy; a day with cloud cover exceeding 70% is considered overcast; and a day with precipitation greater than 0 is considered rainy.

[0050] The Pearson correlation coefficients with the target station were calculated based on the cloud cover time series and precipitation time series, respectively, to obtain the cloud cover correlation coefficient and precipitation correlation coefficient.

[0051] The formula for calculating the Pearson correlation coefficient is:

[0052]

[0053] In the formula, r is the Pearson correlation coefficient; when r=1, it is a positive correlation, indicating that the two stations are highly synchronized in terms of weather changes; when r=0, there is no linear relationship; when r=-1, it is a negative correlation. The weather characteristics of the target station on day i; The weather characteristics of similar stations on day i; This represents the average weather characteristics of the target station over N days. This represents the average weather characteristics of similar stations over N days; N is the total number of days in the historical data.

[0054] By selecting similar meteorological features from the set of similar stations that have a cloud cover correlation coefficient greater than 0.85 and a precipitation correlation coefficient greater than 0.85, a set of similar meteorological features is obtained.

[0055] Furthermore, the historical theoretical power difference and historical similar meteorological characteristics are input into a temporal convolutional neural network to predict the power difference and calculate the non-sunny day power prediction results for the target station, including:

[0056] After aligning the set of similar meteorological features with historical theoretical power difference data, the data is input into a temporal convolutional network to predict the power difference of the target station, including:

[0057] The expression for a temporal convolutional network is:

[0058]

[0059]

[0060] In the formula, denoted as the predicted power difference of the target station at time point s; k is the convolution kernel size. This is the expansion offset; This represents the temporal position in the input sequence that is accessed by the convolutional kernel. ; x is the input after aligning the set of similar meteorological features with the historical theoretical power difference data; To expand the causal convolution kernel; The i-th weight coefficient of the convolution kernel; This refers to the output of a certain layer in a temporal convolutional network. To activate the function, a nonlinear capability is introduced; For residual connections, the original input is added to the result of the convolution transformation, which helps stabilize the training of deep networks;

[0061] The predicted non-clear-sky power value of the target power station is calculated based on the predicted power difference, and the expression is as follows:

[0062]

[0063] In the formula, The non-clear-sky power prediction value of the target station at time point s; The photovoltaic power at time point s in the clear sky photovoltaic curve; This represents the difference in predicted power at the target power station at time point s.

[0064] Furthermore, power prediction is performed on all photovoltaic power stations within the photovoltaic cluster to obtain the predicted power of the photovoltaic cluster, including:

[0065] The expression for predicting the power of a photovoltaic cluster is:

[0066]

[0067] In the formula, This represents the predicted power output of the photovoltaic cluster at time point s. is the power prediction value of the i-th photovoltaic power station at time point s, including the power prediction results for sunny days and the power prediction results for non-sunny days; n is the number of photovoltaic power stations in the photovoltaic cluster.

[0068] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0069] This invention discloses a photovoltaic cluster power prediction method based on the mining of related features of similar power plants. The method analyzes the weather characteristics of adjacent power plants and calculates the historical theoretical power difference of the target power plant. It then applies a temporal convolutional network model to capture long-term dependencies in the sequence, thereby improving the prediction accuracy and reliability. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0072] Figure 2 This is a schematic diagram comparing the clear-sky photovoltaic curves in historical data in this embodiment;

[0073] Figure 3 This is a schematic diagram showing the difference in rainfall, cloud cover, historical theoretical power, and power comparison under different weather types in this embodiment;

[0074] Figure 4 This is a comparison chart of prediction results under different information input modes in non-sunny weather conditions in this embodiment;

[0075] Figure 5 This is a schematic diagram comparing the prediction curves of different methods under sunny weather conditions in this embodiment;

[0076] Figure 6 This is a schematic diagram comparing the prediction curves of different methods under non-sunny weather conditions in this embodiment;

[0077] Figure 7 This is a schematic diagram of the prediction curves for different methods in photovoltaic cluster A in this embodiment;

[0078] Figure 8 This is a schematic diagram of the prediction curves for different methods in photovoltaic cluster B in this embodiment;

[0079] Figure 9 This is a schematic diagram of the prediction curves for different methods in photovoltaic cluster C in this embodiment. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] The purpose of this invention is to provide a photovoltaic cluster power prediction method based on the mining of relevant features of similar power plants, aiming to solve or improve at least one of the above-mentioned technical problems.

[0082] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] like Figure 1 As shown, this invention provides a photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants, including:

[0084] Step 1: Obtain historical photovoltaic power and numerical weather forecast data for the target site. Divide the forecast days into clear weather forecast days and non-clear weather forecast days, including:

[0085] Step 11: Detrend and normalize the historical photovoltaic power and numerical weather forecast data. Normalize the historical photovoltaic power data and numerical weather forecast data to eliminate the impact of differences in the dimensions of different data features on the weights of the prediction model.

[0086] Step 12: Use NWP (Non-Sunny Weather Prediction) detection to classify the predicted days into sunny prediction days and non-sunny prediction days. Determine the weather type for each prediction day based on cloud cover and precipitation thresholds in the NWP: if cloud cover is less than 30% of the sky area and precipitation is 0 during the prediction period, it is considered sunny. If cloud cover is between 30% and 70%, exceeds 70%, or precipitation is greater than 0, it is collectively referred to as non-sunny. Based on this classification, subsequent steps will construct different prediction model paths for sunny and non-sunny prediction days respectively.

[0087] Step 2: For the sunny day prediction day, the pre-trained temporal convolutional network model is used directly for prediction to obtain the sunny day power prediction result;

[0088] The power prediction process for the target power station on the clear weather forecast day is as follows:

[0089] S1. Filter highly correlated meteorological data, select sunny day samples from historical data, and calculate the Pearson correlation coefficient between each meteorological feature and the photovoltaic power sequence. Set the correlation threshold to 0.85, and select meteorological features with correlation coefficients greater than 0.85 as key input variables to construct the input feature set for the prediction model.

[0090] S2 constructs a temporal convolutional network as a predictor. This model employs a dilated causal convolutional structure, expanding the receptive field by setting different dilation coefficients.

[0091] S3 uses the historical clear-day meteorological feature data selected in S1 as input and the corresponding historical actual power data as labels to supervise the training of the model and learn the nonlinear mapping relationship between meteorological factors and photovoltaic power.

[0092] S4 inputs the meteorological data of the day to be predicted into the trained temporal convolutional network model, calculates the normalized predicted power, and finally outputs the clear-day power prediction result of the target station after inverse normalization processing.

[0093] Step 3: For non-sunny forecast days, construct the detrended historical theoretical power difference of the target power station, and at the same time, screen similar power stations based on power generation and weather characteristics, determine historical similar meteorological characteristics, input the historical theoretical power difference and historical similar meteorological characteristics into the temporal convolutional neural network, predict the power difference, and calculate the non-sunny power prediction result of the target power station.

[0094] The power prediction process for the target power station on non-sunny forecast days is as follows:

[0095] S1, Based on the historical photovoltaic power of the target power plant, construct the clear-sky photovoltaic curve for the target power plant and calculate the detrended historical theoretical power difference, including:

[0096] S11, based on a preset search range, search the historical data of the target power plant for the actual power output and construct a clear-sky photovoltaic curve, including:

[0097] The solar photovoltaic curve under clear skies is expressed as follows:

[0098]

[0099] In the formula, Let be the clear-sky power baseline value at time point i in the k-th search interval, representing the historical maximum power output at time point i within the search interval; Let m be the actual power value at time point i on day m in the k-th search interval; m is the number of days in each search interval, and in this embodiment m=15;

[0100] In the above steps, the historical data includes the target site's historical data for the entire year, up to 10 days before the forecast date; the clear sky photovoltaic curve shows how the time of the first and last non-zero power point of the photovoltaic cluster in a day changes as the months progress throughout the year.

[0101] S12, based on the clear-sky photovoltaic curve, calculate the historical theoretical power difference of the photovoltaic cluster detrending, including:

[0102] Based on the clear-sky photovoltaic curve, the historical theoretical power difference of the target power station that has not detrended is calculated, and the expression is:

[0103]

[0104] In the formula, The undetrended historical theoretical power difference is the value of time point i on day m in the k-th search interval; Let i be the daily net air power value at time point i in the k-th search interval; The actual power value at time point i of the power data on day m in the k-th search interval;

[0105] Based on the undetrended historical theoretical power difference, the detrended historical theoretical power difference is calculated, expressed as:

[0106]

[0107] In the formula, It represents the detrended historical theoretical power difference of time point i on day m in the k-th search interval.

[0108] S2, searches for a set of similar stations to the target station based on dynamic time-warped distance, including:

[0109] S21, Construct the distance matrix M between the target station and the reference station, including:

[0110] Given that the target site has m time points, the first time series is: The reference station has n time points, resulting in the second time series. Calculate the Euclidean distance between different time points and generate the distance matrix M, expressed as:

[0111]

[0112]

[0113] In the formula, The distance is Euclidean. The power generation at the i-th time point in the first time series; The power generation at the j-th time point in the second time series;

[0114] S22, calculate the minimum cumulative distance based on the distance matrix M, and generate the cumulative distance matrix W, expressed as:

[0115]

[0116]

[0117] In the formula, The minimum cumulative distance required to walk from the starting point to position (i,j); The Euclidean distance between the two time points; The goal is to select the position with the lowest cumulative cost from the three front-end positions.

[0118] Constraints are applied to the above cumulative calculations, including:

[0119] Boundary constraints are applied, and the path calculation follows the time series order, proceeding sequentially from the lower left corner to the upper right corner of the distance matrix M.

[0120] The continuity constraint stipulates that a point cannot be matched with any other point, but can only be matched with the point before or after it to ensure that each coordinate is in the cumulative distance matrix W.

[0121] Monotonicity constraint: the elements in the cumulative distance matrix W must be monotonic over time during computation.

[0122] S23, Based on the cumulative distance matrix W, the DTW distance between the target station and the reference station is obtained, expressed as:

[0123]

[0124] In the formula, The DTW distance between the first and second time series is represented by a smaller value, indicating that the invention modes of the target site and the reference site are more similar. The value at the top right corner of the cumulative distance matrix W represents the total cost of optimal alignment from beginning to end.

[0125] S24, DTW distance between the target station and the reference station, generating DTW vectors with multiple reference stations, expressed as:

[0126]

[0127] In the formula, Let be the DTW distance between the i-th target station and the n-th reference station; The time series of the i-th target station; This is the time series of the nth reference station;

[0128] S25, select the k most similar reference stations to obtain the set of similar stations, expressed as:

[0129]

[0130] In the formula, Let i be the set of similar stations to the i-th target station; Let k be the kth similar station with the smallest DTW distance; k is the number of similar stations.

[0131] S3, based on cloud cover and precipitation, historical meteorological characteristics of similar stations are filtered to obtain a set of similar meteorological characteristics, including:

[0132] S31. Based on cloud cover and precipitation, historical data from similar stations are classified for weather conditions to obtain cloud cover time series and precipitation time series.

[0133] A day with cloud cover less than 30% of the sky is considered sunny; a day with cloud cover between 30% and 70% is considered cloudy; a day with cloud cover exceeding 70% is considered overcast; and a day with precipitation greater than 0 is considered rainy.

[0134] S32, calculate the Pearson correlation coefficient with the target station based on the cloud cover time series and precipitation time series respectively, and obtain the cloud cover correlation coefficient and precipitation correlation coefficient;

[0135] The formula for calculating the Pearson correlation coefficient is:

[0136]

[0137] In the formula, r is the Pearson correlation coefficient; when r=1, it is a positive correlation, indicating that the two stations are highly synchronized in terms of weather changes; when r=0, there is no linear relationship; when r=-1, it is a negative correlation. The weather characteristics of the target station on day i; The weather characteristics of similar stations on day i; This represents the average weather characteristics of the target station over N days. This represents the average weather characteristics of similar stations over N days; N is the total number of days in the historical data.

[0138] S33. Select meteorological features from historical data of similar stations that have a correlation coefficient of cloud cover greater than 0.85 and a correlation coefficient of precipitation greater than 0.85 to obtain a set of similar meteorological features.

[0139] S4 aligns the set of similar meteorological features with historical theoretical power difference data, then inputs it into a temporal convolutional network to predict the power difference of the target station, including:

[0140] The expression for a temporal convolutional network is:

[0141]

[0142]

[0143] In the formula, denoted as the predicted power difference of the target station at time point s; k is the convolution kernel size. This is the expansion offset; This represents the temporal position in the input sequence that is accessed by the convolutional kernel. ; x is the input after aligning the set of similar meteorological features with the historical theoretical power difference data; To expand the causal convolution kernel; The i-th weight coefficient of the convolution kernel; This refers to the output of a certain layer in a temporal convolutional network. To activate the function, a nonlinear capability is introduced; For residual connections, the original input is added to the result of the convolution transformation, which helps stabilize the training of deep networks;

[0144] The predicted non-clear-sky power value of the target power station is calculated based on the predicted power difference, and the expression is as follows:

[0145]

[0146] In the formula, The non-clear-sky power prediction value of the target station at time point s; The photovoltaic power at time point s in the clear sky photovoltaic curve; The predicted power difference of the target power station at time point s;

[0147] The expression for predicting the power of a photovoltaic cluster is:

[0148]

[0149] In the formula, This represents the predicted power output of the photovoltaic cluster at time point s. is the power prediction value of the i-th photovoltaic power station at time point s, including the power prediction results for sunny days and the power prediction results for non-sunny days; n is the number of photovoltaic power stations in the photovoltaic cluster. Specific Implementation

[0151] In this embodiment, the power of the three photovoltaic clusters at stations A, B, and C is predicted in the short term.

[0152] The data used includes measured power data from the aforementioned stations and corresponding NWP meteorological data (including irradiance, temperature, precipitation, cloud cover, humidity, wind speed, wind direction, and atmospheric pressure), with a time resolution of 15 minutes. The training and test sets are divided in a 1:1 ratio.

[0153] Step 1: Data Preprocessing and Weather Classification. Taking photovoltaic power station A as an example, 2017 data is used to extract the clear-sky power curve and the historical theoretical power difference, while 2018 data is used for evaluation. All raw data are first normalized using the following expression:

[0154]

[0155] In the formula, It is the original time series data. It is a sequence The minimum value, It is a sequence The maximum value, It is normalized data.

[0156] Subsequently, the weather for the forecast day was classified based on cloud cover and precipitation thresholds from numerical weather prediction data. Cloud cover below 30% was defined as a sunny day, while cloud cover above 70% or precipitation of non-zero was defined as a cloudy day.

[0157] like Figure 2 As shown in the figure, Examples 1 to 10 represent historical measured power curves for different dates within the search interval. Power curves differ depending on the weather conditions; by extracting the maximum value at each moment, the clear-sky photovoltaic curve for that period can be fitted.

[0158] like Figure 3 As shown, the historical theoretical power difference for predicting target power stations under non-sunny weather scenarios is constructed. In non-sunny weather scenarios, the historical theoretical power difference serves as the new prediction target. Daily clear-sky power curves are extracted, and by searching historical power data, the historical maximum power value at each time point is extracted, and a smooth, ideal clear-sky curve is fitted. Then, the theoretical power difference is calculated by subtracting the actual power under non-sunny weather scenarios from the clear-sky curve, followed by detrending processing, to obtain the historical theoretical power difference sequence. Validity analysis of the historical theoretical power difference is then performed.

[0159] As shown in Table 1, under severe weather conditions, the correlation between traditional power and irradiance decreases due to the inaccuracy of NWP; while the historical theoretical power difference constructed in this invention shows a significantly improved physical correlation with cloud cover and precipitation, proving that the historical theoretical power difference is a more reasonable and stable prediction target.

[0160]

[0161] like Figure 3As shown in Table 1, the comparison of power, historical theoretical power difference, and meteorological data under non-sunny weather conditions is presented in the form of normalized values. Table 1 shows that under sunny weather conditions, there is a high correlation between irradiance, temperature, humidity, and power. The temporal correlation between precipitation and cloud cover, as well as atmospheric pressure and power, is negligible. Therefore, traditional modeling methods can still be used in this invention. Under non-sunny weather conditions, the correlation between power and key meteorological data decreases. However, irradiance still maintains a relatively high correlation with power. It should be noted that this high correlation is mainly due to the inherent daily periodicity of photovoltaic systems. Subsequence experiments revealed that traditional modeling methods still lead to underfitting due to this characteristic. Rainfall and cloud cover are important meteorological characteristics affecting photovoltaic power generation at the physical level. Under sunny weather conditions, the correlation between the theoretical power difference and most meteorological elements is relatively low. However, under extreme weather conditions, the correlation between the theoretical power difference and cloud cover and precipitation increases, and cannot be simply attributed to a linear relationship. Increased cloud cover or rainfall leads to varying degrees of reduction in photovoltaic output, thereby increasing the theoretical power difference.

[0162] Step 3: Similar Station Selection and Feature Mining Based on Dynamic Time Warping Distance. Searching for Similar Neighboring Stations: Using the dynamic time warping algorithm, the minimum cumulative distance is calculated by comparing the historical power sequences of the target station and neighboring stations, selecting the K neighboring stations with the most similar power fluctuation patterns. Filtering Effective Information: Among the similar neighboring stations selected by DTW, the similarity between their meteorological characteristics and the target station is calculated. Features with a correlation coefficient greater than 0.85 are selected to form the optimal input information set.

[0163] Step 4: Verify the predictive performance of information filtering under non-clear weather conditions by establishing two comparison methods. Method 1: Introduce multi-source meteorological information from neighboring locations filtered by DTW as input. Method 2: Use only single meteorological information from the target station as input. For example... Figure 4 As shown in Table 2, using filtered information can improve the accuracy and stability of predictions. However, it must be recognized that not all stations have sufficient meteorological data to support clustering modeling methods, and this geographical constraint is also a factor limiting the stability of modeling. This limitation is particularly pronounced for photovoltaic stations that rely on the same meteorological station.

[0164] Table 2. Predictive performance of non-clear weather conditions under different information input modes.

[0165]

[0166] The expression for the error index is:

[0167]

[0168]

[0169]

[0170] In the formula, This refers to the installed capacity. It represents the number of test samples. It is a standardized actual power value. It is a standardized predicted power value.

[0171] Step 5: Establish different prediction models and compare them with existing methods to verify the overall effectiveness of the method. Method 1: Predict the historical theoretical power difference using a time-convolutional network model featuring precipitation and cloud cover. Method 2: Predict the power using a time-convolutional network model featuring irradiance and temperature. Method 3: Predict the historical theoretical power difference using a long short-term memory network model featuring precipitation and cloud cover. Method 4: Predict the power using a long short-term memory network model featuring irradiance and temperature. Method 5: Autoregressive moving average modeling method. Method 6: Predict the power using a multiple linear regression modeling method featuring irradiance and temperature. Method 7: Predict the power using a time-series ensemble model. Method 8: Predict the power using a combined prediction method based on multimodal decomposition, parallel bidirectional long short-term memory networks, and convolutional neural networks. (From Table 3 and...) Figure 5 It can be seen that, compared with traditional methods, the method for predicting the difference in historical theoretical power under clear sky conditions has limited practical value, with the latter being significantly superior to the former. Compared with cloudy and rainy weather conditions, clear sky conditions exhibit less interference from rainfall and cloud cover on photovoltaic output, and these characteristics are considered redundant factors when used for modeling.

[0172] Table 3 Error Evaluation Table for Each Prediction Method

[0173]

[0174] Under severe weather conditions, the method for predicting historical theoretical power differences shows varying degrees of improvement across different types of cloudy and rainy weather. Compared to other traditional modeling methods and several recently proposed approaches, the method presented in this paper can more accurately track overall trend changes, validating the effectiveness of the entire prediction process. However, even using the proposed method, it still cannot accurately track power fluctuations at every time point. This limitation stems from the fact that precipitation and cloud cover forecasts in traditional numerical weather prediction are specific to certain regions, leading to reduced accuracy in short-term forecasts and failing to guarantee accurate forecasting at every moment.

[0175] Step 6: Establish a photovoltaic power prediction model considering data from multiple clusters. This includes Method 1: a temporal convolutional network modeling method without filtering information, using precipitation and cloud cover as features to predict the historical theoretical power difference. Method 2: a temporal convolutional network modeling method using irradiance and temperature as features to predict power. Method 3: the method proposed in this paper. Photovoltaic clusters exhibit different meteorological conditions due to their varying distribution ranges. However, overall, the number of sunny days exceeds the number of hazardous weather types. As a result, the power periodicity of the clusters is more pronounced than the daily periodicity of individual stations.

[0176] Therefore, using only historical theoretical power difference prediction methods cannot achieve optimal prediction results. For example... Figures 7-9 As shown in Table 4, a certain degree of overfitting was observed in the daily forecast performance. Compared with traditional power forecasting methods, the proposed method has a stronger ability to detect power fluctuations. By combining the advantages of traditional and novel methods, the proposed method achieves optimal performance across various indicators and obtains the most accurate forecast results by utilizing the clustering effect. Finally, through case studies of three photovoltaic cluster sites, it was found that the proposed method has strong robustness and stability, and exhibits good forecasting capabilities at sites with different geographical environments.

[0177] Table 4

[0178]

[0179] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0180] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants, characterized in that, include: The historical photovoltaic power and numerical weather forecast data of the target site are obtained, and the forecast days are divided into clear weather forecast days and non-clear weather forecast days. For the sunny day prediction, the pre-trained temporal convolutional network model is used directly to make the prediction and obtain the sunny day power prediction result. For non-sunny forecast days, the historical theoretical power difference of the target power station is detrended. At the same time, similar power stations are selected based on power generation and weather characteristics to determine historical similar meteorological characteristics. The historical theoretical power difference and historical similar meteorological characteristics are input into a temporal convolutional neural network to predict the power difference and calculate the non-sunny power prediction result of the target power station. Power prediction is performed on all photovoltaic power stations in the photovoltaic cluster to obtain the predicted power of the photovoltaic cluster; The construction of the historical theoretical power difference for detrending the target power station includes: The real power in the historical data of the target power station is searched according to the preset search range to construct the clear sky photovoltaic curve; Based on the clear-sky photovoltaic curve, calculate the historical theoretical power difference of the photovoltaic cluster detrending, specifically including: Based on the clear-sky photovoltaic curve, the historical theoretical power difference of the target power station that has not detrended is calculated, and the expression is: In the formula, The undetrended historical theoretical power difference is the value of time point i on day m in the k-th search interval; Let i be the daily net air power value at time point i in the k-th search interval; The actual power value at time point i of the power data on day m in the k-th search interval; Based on the undetrended historical theoretical power difference, the detrended historical theoretical power difference is calculated, expressed as: In the formula, The detrended historical theoretical power difference is the value of time point i on day m in the k-th search interval; The step of simultaneously screening similar power stations based on power generation capacity and weather characteristics to determine a set of similar meteorological characteristics includes: Search for a set of similar stations to the target station based on dynamic time-normalized distance; Based on cloud cover and precipitation, historical weather characteristics of similar stations were filtered to obtain a set of similar meteorological characteristics, specifically including: Weather classification was performed on historical data from similar stations based on cloud cover and precipitation to obtain cloud cover time series and precipitation time series; A day with cloud cover less than 30% of the sky is considered sunny; a day with cloud cover between 30% and 70% is considered cloudy; a day with cloud cover exceeding 70% is considered overcast; and a day with precipitation greater than 0 is considered rainy. The Pearson correlation coefficients with the target station were calculated based on the cloud cover time series and precipitation time series, respectively, to obtain the cloud cover correlation coefficient and precipitation correlation coefficient. The formula for calculating the Pearson correlation coefficient is: In the formula, r is the Pearson correlation coefficient; when r=1, it is a positive correlation, indicating that the two stations are highly synchronized in terms of weather changes; when r=0, there is no linear relationship; when r=-1, it is a negative correlation. The weather characteristics of the target station on day i; The weather characteristics of similar stations on day i; This represents the average weather characteristics of the target station over N days. This represents the average weather characteristics of similar stations over N days; N is the total number of days in the historical data. By selecting similar meteorological features from the set of similar stations that have a cloud cover correlation coefficient greater than 0.85 and a precipitation correlation coefficient greater than 0.85, a set of similar meteorological features is obtained.

2. The photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants according to claim 1, characterized in that, The step of searching for the actual power in the historical data of the target power station according to a preset search interval to construct a clear-sky photovoltaic curve includes: The solar photovoltaic curve under clear skies is expressed as follows: In the formula, Let be the clear sky power value at time point i in the k-th search interval; Let be the actual power value at time point i of the power data on day m in the k-th search interval; m is the number of days in each search interval.

3. The photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants according to claim 1, characterized in that, The set of similar stations for the target station searched based on dynamic time-warped distance includes: Construct the distance matrix M between the target station and the reference station, including: Given that the target site has m time points, the first time series is: The reference station has n time points, resulting in the second time series. Calculate the Euclidean distance between different time points and generate the distance matrix M, expressed as: In the formula, The distance is Euclidean. The power generation at the i-th time point in the first time series; The power generation at the j-th time point in the second time series; Calculate the minimum cumulative distance based on the distance matrix M, and generate the cumulative distance matrix W, expressed as: In the formula, The minimum cumulative distance required to walk from the starting point to position (i,j); The Euclidean distance between the two time points; Based on the cumulative distance matrix W, the DTW distance between the target station and the reference station is obtained, expressed as: In the formula, The DTW distance between the first and second time series is represented by a smaller value, indicating that the invention modes of the target site and the reference site are more similar. The value at the top right corner of the cumulative distance matrix W represents the total cost of optimal alignment from beginning to end; The DTW distance between the target station and the reference stations is used to generate DTW vectors with multiple reference stations, expressed as follows: In the formula, Let be the DTW distance between the i-th target station and the n-th reference station; The time series of the i-th target station; This is the time series of the nth reference station; Selecting the k most similar reference stations yields the set of similar stations, expressed as: In the formula, Let i be the set of similar stations to the i-th target station; Let k be the kth similar station with the smallest DTW distance; k is the number of similar stations.

4. The photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants according to claim 1, characterized in that, The process of inputting historical theoretical power differences and historical similar meteorological characteristics into a temporal convolutional neural network to predict power differences and calculate the non-sunny day power prediction results for the target station includes: After aligning the set of similar meteorological features with historical theoretical power difference data, the data is input into a temporal convolutional network to predict the power difference of the target station, including: The expression for a temporal convolutional network is: In the formula, denoted as the predicted power difference of the target station at time point s; k is the convolution kernel size. This is the expansion offset; This represents the temporal position in the input sequence that is accessed by the convolutional kernel. ; x is the input after aligning the set of similar meteorological features with the historical theoretical power difference data; To expand the causal convolution kernel; The i-th weight coefficient of the convolution kernel; This refers to the output of a certain layer in a temporal convolutional network. To activate the function, a nonlinear capability is introduced; For residual connections, the original input is added to the result of the convolution transformation, which helps stabilize the training of deep networks; The predicted non-clear-sky power value of the target power station is calculated based on the predicted power difference, and the expression is as follows: In the formula, The non-clear-sky power prediction value of the target station at time point s; The photovoltaic power at time point s in the clear sky photovoltaic curve; This represents the difference in predicted power at the target power station at time point s.

5. The photovoltaic cluster power prediction method based on the mining of correlation features of similar power plants according to claim 1, characterized in that, The process of predicting the power of all photovoltaic power stations in a photovoltaic cluster to obtain the predicted power of the photovoltaic cluster includes: The expression for predicting the power of a photovoltaic cluster is: In the formula, This represents the predicted power output of the photovoltaic cluster at time point s. is the power prediction value of the i-th photovoltaic power station at time point s, including the power prediction results for sunny days and the power prediction results for non-sunny days; n is the number of photovoltaic power stations in the photovoltaic cluster.

Images