A refined load forecasting method and system for active distribution networks

Abstract

The present invention relates to a refined load forecasting method and system for active power distribution networks. The method accurately divides photovoltaic power and electricity load, cleans the photovoltaic power and electricity load data using different methods based on the type of bad data, builds a model using the convolutional neural network ResNet18_1D to predict photovoltaic power, explores the load characteristics of various industries, and builds a model using a recurrent neural network GRU structure to predict electricity load. The photovoltaic power and electricity load prediction values are accumulated to obtain a refined load forecast result for the active power distribution network. This method can achieve accurate prediction of active power load, meets the requirements of load forecasting for new active power distribution networks, and improves prediction accuracy.

Description

Technical Field

[0001] The present invention relates to a refined load forecasting method and system for an active power distribution network. Background Art

[0002] With the large-scale integration of distributed generation, the traditional distribution network operation mode has changed, potentially leading to power flow reversal, voltage over-limit, and relay protection misoperation. Currently, there is a lack of effective evaluation methods and operational optimization support for the operation of distributed power generation integration. The distribution network's perception and control capabilities are insufficient, the coverage rate of distribution terminals is low, and the grid information collection capabilities are insufficient. The level of data access, governance, and analysis and application still needs to be improved. Moreover, the integration of a large number of distributed power generation has changed the shape and characteristics of the load curve, further reducing the versatility of the load model and making it unable to adapt to future grid development. Most existing distribution network load forecasting methods only consider equivalent load data and rely solely on historical load data as the data source to build models. They do not consider the impact of distributed power generation integration on the shape and characteristics of the load curve, do not accurately divide the generation and consumption components of the active distribution network load, and do not consider the influence of factors such as load industry type and meteorological environment, which affects the accuracy of load forecasting.

[0003] PV power forecasting can be categorized into physical, statistical, machine learning, and hybrid methods. Physical methods are mathematical models based on the principles of PV power generation. They utilize data such as solar irradiance, temperature, humidity, cloud cover, air pressure, wind speed, and wind direction obtained through numerical weather forecasting to directly calculate PV power generation. Statistical methods use processed historical data such as solar irradiance / wind speed and PV power generation output to establish a correlation mapping relationship (i.e., a data model) between input and output data through curve fitting, parameter estimation, and correlation analysis. This allows for predictions of future PV power generation output. Common statistical forecasting methods include time series analysis, fuzzy theory, grey theory, regression analysis, and spatiotemporal correlation. Machine learning methods usually refer to the use of neural networks to learn the mapping relationship between input and output through relevant algorithms to make predictions, and have become a relatively common method for photovoltaic predictions. The commonly used methods for electricity load forecasting in China are mainly artificial intelligence prediction methods, such as support vector machines, cloud computing methods combined with multivariate regression methods, convolutional neural networks and long and short-term memory artificial neural networks, XGBoost algorithms, long and short-term memory artificial neural networks based on attention mechanisms, GRU-NN models, etc., which have reduced load forecasting errors to a certain extent. Summary of the Invention

[0004] The purpose of the present invention is to provide a refined load forecasting method and system for active distribution networks, accurately divide photovoltaic power generation components and electricity consumption components, explore the load characteristics of various industries, predict photovoltaic power and electricity load respectively, and finally accumulate and obtain refined load forecast results for active distribution networks.

[0005] To achieve the above object, the technical solution of the present invention is: a refined load forecasting method for active distribution network, comprising the following steps:

[0006] Step S1, load / power data cleaning: dividing the distribution transformer equivalent load into the power generation component and the power load component of the distributed photovoltaic power data according to the topological structure and the distributed photovoltaic power data; and performing data cleaning on the distributed photovoltaic power data and the power load data respectively;

[0007] Step S2, distributed photovoltaic power prediction: A model is established for each distributed photovoltaic station, and short-term and ultra-short-term power predictions are performed based on the model. Then, based on the topological relationship and geographical location, photovoltaic power prediction results are accumulated at the distribution transformer / substation level, feeder level, substation level, district / county level, and prefecture / city level.

[0008] Step S3, industry load characteristics analysis: The industry load characteristics analysis includes two parts. The first step is to analyze the relationship between the load of each industry and different meteorological factors through historical load data, and to explore the main influencing factors of each industry. The second step is to use cluster analysis methods to extract typical characteristic curves of similar load days.

[0009] Step S4, refined load forecasting: Refined load forecasting includes three steps: ① Establish a load forecasting model for each distribution transformer based on the results of industry characteristic analysis; ② Obtain the power load forecast results of the distribution transformer based on the model, and combine them with the photovoltaic power forecast results under the corresponding distribution transformer to accumulate the refined load forecast results of the corresponding distribution transformer; ③ Based on the topological relationship and geographical location, the load forecast results of each level of the city, district, county, substation, and feeder are obtained based on the forecast results of the distribution transformer.

[0010] In one embodiment of the present invention, step S1 is specifically implemented as follows:

[0011] (1) Distributed photovoltaic power data cleaning

[0012] There are three types of abnormal data for photovoltaic power generation: locally missing values, abnormal fluctuation data, and non-normal data. The methods for handling these three types of bad data are as follows:

[0013] ① Local missing values: For missing values, dynamic weight correction is used to compensate for the missing values at the current moment based on the data of the same weather and time in the previous four days. If there is a lot of missing data in a day, the data of the entire day is directly deleted. The dynamic weight correction method is as follows:

[0014]

[0015] y'1 represents missing values, is the measured value at the same time in the previous n days with the same weather as the missing value; η i represents random weights, i = 1, 2, 3, 4, and η1 + η2 + η3 + η4 = 1;

[0016] ② Abnormal fluctuation data: Abnormal fluctuation data of photovoltaic power generation power can be divided into two situations: one is the fluctuation of photovoltaic power generation power due to sudden changes in weather conditions; the other is the abnormal fluctuation of photovoltaic power data despite no obvious fluctuation in weather conditions. The former is a normal phenomenon, while the latter is bad data and needs to be eliminated. When abnormal fluctuation data is detected, it is necessary to first check whether there is a significant change in the meteorological data. If there is a significant change, the data is skipped. Otherwise, the data is eliminated and the dynamic weight correction method is used to re-correct the data.

[0017] ③ For abnormal data, if the data exceeds the normal range and exceeds the maximum power generation of the photovoltaic power station, it will be treated as abnormal fluctuation data. If a negative number appears, it will be replaced by 0;

[0018] (2) Power load data cleaning

[0019] Bad load data can be categorized into three types: locally missing values, abnormally fluctuating data, and abnormal data. The processing methods are as follows:

[0020] For abnormal load fluctuation data, the density clustering algorithm DBSCAN is used to identify abnormal fluctuation data in load history data. During the model training process of the DBSCAN algorithm, there are two important parameters that need to be clarified. One parameter is the minimum sample value min_samples, which represents the minimum number of sample points to form a cluster; the other parameter is the distance eps, which represents the maximum distance between sample points in the same cluster. When a sample point is not within the cluster circle with eps as the unit, this sample point is likely to be an anomaly. The minimum sample value min_samples for load data cleaning is set to 10% of the number of samples for a single identification, and the distance eps is set to 10 times the standard deviation of the sample.

[0021] According to the actual situation of the load history data to be corrected, when there are no abnormal points in the load data of the same type of days on the recent working days or holidays on the day of data cleaning, the daily load of the same type of days is used as the basic curve and the original load data is corrected using the multiple ratio grafting method. Otherwise, the linear interpolation method is used for correction; the formula of the multiple ratio grafting method is as follows:

[0022]

[0023] i=p,p+1,p+2,…,q

[0024] Where x c is the corrected value of the abnormal data point, x t is the value of the base curve, x d is the value of the correct point of the curve to be corrected, and p to q are abnormal data points;

[0025] Linear interpolation is to use a straight line passing through the points (x1, y1) and (x2, y2) to approximate the function The specific formula is as follows:

[0026]

[0027] In one embodiment of the present invention, step S2 is specifically implemented as follows:

[0028] 1) Short-term power forecast

[0029] A distributed photovoltaic short-term power prediction model was built using the convolutional neural network ResNet18_1D structure. Meteorological data including historical photovoltaic power, irradiance, and temperature were used as input features, and the predicted daily photovoltaic power was used as the training target to build a short-term power prediction model for distributed photovoltaic short-term power prediction. Based on the grid meteorological data input model provided by numerical weather forecasting, the distributed photovoltaic short-term power prediction results for the next three days were obtained. Based on the input-output model establishment method with limited input delay and the structural characteristics of the convolutional neural network, a one-dimensional convolutional neural network prediction model was designed. Multiple data from each day were selected to construct a one-dimensional convolutional neural network input-output model for photovoltaic power prediction, and the time correlation between similar times was extracted. Where X = {x1, x2,…, x 96}, the number of input variables for a single day is 96, Y is the photovoltaic power, and the predicted weather forecast data corresponds to the photovoltaic power of the day;

[0030] The model design and training parameter design are as follows:

[0031] Convolutional layer design: the moving stride is set to 1, and the boundary filling method is selected as SAME;

[0032] Pooling layer design: Select the maximum pooling method; taking into account the asymmetry of the length and width of the input data, use a mixture of 1*2 and 2*2 maximum pooling convolution kernels with a stride of 2;

[0033] Activation function: Select LeakyRelu function;

[0034] Fully connected layer design: Keep the dimension consistent with the output Y;

[0035] Network optimization training parameter design: exponential decay gradient descent is selected as the gradient descent method, Adam method is selected as the optimization algorithm, and SmoothL1Loss is calculated using the power prediction value and the actual value as the loss function. At the same time, to prevent overfitting, a dropout layer is added to the model.

[0036] Data partitioning: The data partitioning standard is to divide historical data into training data and test data in a ratio of 8:2. All test data is used as validation data for model validation during training, and then input into the network for network training, validation and testing;

[0037] Learning rate decay strategy: using exponential decay;

[0038] 2) Ultra-short-term power forecast

[0039] A recurrent neural network algorithm is used to train an ultra-short-term irradiance prediction model based on data collected by meteorological monitoring devices. The extreme gradient boosting algorithm XgBoost is used to train an ultra-short-term photovoltaic power prediction model. For distributed photovoltaics within the range of the meteorological device, the ultra-short-term irradiance data and the daily data including temperature and humidity obtained by edge calculation of the meteorological monitoring device are input into the ultra-short-term photovoltaic power prediction model for prediction. For distributed photovoltaics outside the range of the meteorological device, the ultra-short-term irradiance prediction data is obtained by grid meteorological data + meteorological monitoring device correction, and then input into the ultra-short-term photovoltaic power prediction model for prediction, and the distributed photovoltaic ultra-short-term power prediction results for the next 4 hours are output.

[0040] In one embodiment of the present invention, step S3 is specifically implemented as follows:

[0041] 1) Analysis of influencing factors

[0042] The correlation analysis method is used to obtain the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall, and the main influencing factors of the load of different industries are obtained as a reference for the load forecasting model structure; the correlation coefficient is used as a reference standard for the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall. The calculation formula is shown below:

[0043]

[0044] Where: R is the correlation coefficient; cov(x, y) is the covariance of x and y; D(x) and D(y) are the mean square errors between the two factors respectively; the correlation coefficient r value ranges from -1 to 1. If r>0, it means that the two factors are positively correlated; if it is between 0 and 1, it means that the two factors are negatively correlated; the closer the absolute value of r is to 1, the stronger the linear relationship between factors x and y is; the closer the absolute value of r is to 0, the weaker the linear relationship between x and y is; 0<|r|<1, it means that x and y are correlated, but the relationship is nonlinear; |r|≥0.7 is considered highly correlated; 0.4<|r|<0.7 is considered significantly correlated; |r|<0.4 is considered weakly correlated;

[0045] 2) Extract typical characteristic curves

[0046] The k-means clustering algorithm with adaptive k value is used to perform cluster analysis on the historical load of each load in the past year to calculate the typical characteristic curve; the cosine similarity distance is used as the cluster distance indicator, and the cosine similarity calculation method is as follows:

[0047]

[0048] Where, P i =[p i1 ,p i2 ,…,p iN ] and P j =[p j1 ,p j2 ,…,p jN ] is the daily load curve of two days i and j, and N is the number of sampling points per day;

[0049] The silhouette coefficient is selected as the evaluation index of the clustering result. The k value is scanned within the predetermined range, and the largest silhouette coefficient is selected as the optimal k value. The clustering result at this time is optimal. The calculation method of the silhouette coefficient is as follows:

[0050]

[0051] Where: a(i) is the sample P i To other samples P in the same cluster j The average distance of sample P; b(i) is the average distance of sample P i To all samples P of other clusters j The average distance of clustering; M is the number of samples, and the silhouette coefficients of all samples are averaged to obtain the silhouette coefficient S of the overall clustering result.

[0052] In one embodiment of the present invention, the specific implementation steps of step S4 are as follows:

[0053] 1) Establish a load forecasting model

[0054] A short-term load forecasting model is built using a recurrent neural network GRU gate unit. The data required for model training includes historical load data of the distribution transformer, distribution transformer ledger data, and historical meteorological data. The short-term load forecasting model based on the recurrent neural network GRU gate unit consists of three GRU layers. A Dropout layer is added after each GRU layer to avoid overfitting during model training. The two fully connected layers in the output layer are used to map the features of the GRU gate unit into 96 points of daily load data for the output layer's 1×96 prediction output. The model training objective is to predict the daily active power. The model's input features include:

[0055] a) The historical load curves of the three days before the forecast date and the typical load curves of similar days, with the historical load data normalized;

[0056] b) Valid meteorological data for the three days before the forecast date and similar days, with normalized meteorological data;

[0057] c) Date type: one-hot encoding is used to process date types;

[0058] 2) Calculate the distribution transformer refined load forecast results

[0059] The historical loads of the three days before the forecast date and the same day last week, as well as the meteorological data and date type of the forecast date are normalized and one-hot encoded before being input into the model to obtain the short-term electricity load forecast results for the forecast date and the next two days, a total of three days.

[0060] The ultra-short-term forecast results for the next 0-4 hours are given by correcting the real-time load with the short-term forecast results:

[0061] The sampling time of the field data points is 15 minutes, so let δt = 15 minutes;

[0062] The ultra-short-term prediction time range is 4 hours, and N = 16;

[0063] The relative error α of short-term load forecasting from the current time t to N δt times before s To predict the relative error α of ultra-short-term prediction us , the ultra-short-term prediction result F at time t+δt us The calculation formula for (t+δt) is as follows:

[0064]

[0065] Where Y(t) and F s (t) represent the actual and short-term forecast values at time t, respectively, W n represents the weight at time tn·δt;

[0066] Add the electricity load forecast results and the photovoltaic power forecast results to obtain the short-term refined load forecast results of the distribution transformer;

[0067] 3) Calculate the load forecast results at each level

[0068] The short-term refined load forecast results of the distribution transformer are predicted, and the refined load forecast results of the equipment at each level of city, district, county, substation and feeder are accumulated according to the topological relationship.

[0069] The present invention also provides a refined load forecasting system for active distribution networks, comprising a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps described above can be implemented.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] (1) Data cleaning is performed on photovoltaic power and electricity load respectively to identify abnormal data, fill in missing data and abnormal data, and prevent abnormal data from affecting model training and prediction results.

[0072] (2) Analyze the relationship between electricity load in different industries and different meteorological factors, explore the main influencing factors of each industry, use cluster analysis method to extract the typical characteristic curve of similar load days, and use it as an important feature of load forecasting to input it into the forecasting model, thus avoiding the influence of photovoltaic power on the stability of load characteristics.

[0073] (3) The photovoltaic power and electricity load of the distribution transformer are predicted separately, and then the results are accumulated to obtain the overall refined load forecast results. The data processing is more refined and the prediction accuracy is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Schematic diagram of load / power data cleaning.

[0075] Figure 2 This is the flow chart of the DBSCAN algorithm.

[0076] Figure 3 This is a structural diagram of the one-dimensional convolutional neural network prediction model.

[0077] Figure 4 This is a flow chart for refined load forecasting.

[0078] Figure 5 Build a short-term load forecasting model structure for the recurrent neural network GRU gate unit. DETAILED DESCRIPTION

[0079] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0080] The present invention provides a refined load forecasting method for an active power distribution network, comprising the following steps:

[0081] Step S1, load / power data cleaning: dividing the distribution transformer equivalent load into the power generation component and the power load component of the distributed photovoltaic power data according to the topological structure and the distributed photovoltaic power data; and performing data cleaning on the distributed photovoltaic power data and the power load data respectively;

[0082] Step S2, distributed photovoltaic power prediction: A model is established for each distributed photovoltaic station, and short-term and ultra-short-term power predictions are performed based on the model. Then, based on the topological relationship and geographical location, photovoltaic power prediction results are accumulated at the distribution transformer / substation level, feeder level, substation level, district / county level, and prefecture / city level.

[0083] Step S3, industry load characteristics analysis: The industry load characteristics analysis includes two parts. The first step is to analyze the relationship between the load of each industry and different meteorological factors through historical load data, and to explore the main influencing factors of each industry. The second step is to use cluster analysis methods to extract typical characteristic curves of similar load days.

[0084] Step S4, refined load forecasting: Refined load forecasting includes three steps: ① Establish a load forecasting model for each distribution transformer based on the results of industry characteristic analysis; ② Obtain the power load forecast results of the distribution transformer based on the model, and combine them with the photovoltaic power forecast results under the corresponding distribution transformer to accumulate the refined load forecast results of the corresponding distribution transformer; ③ Based on the topological relationship and geographical location, the load forecast results of each level of the city, district, county, substation, and feeder are obtained based on the forecast results of the distribution transformer.

[0085] like Figure 1-5 As shown, the present invention provides a refined load forecasting method for an active distribution network with a large number of distributed photovoltaic accesses, specifically comprising the following steps:

[0086] (1) Load / power data cleaning

[0087] like Figure 1 As shown in the figure, the distribution transformer equivalent load is divided into the power generation component and the power load component of the distributed photovoltaic power data according to the topological structure and distributed photovoltaic power data. The distributed photovoltaic power and load data are cleaned separately.

[0088] (1) Distributed photovoltaic power data cleaning

[0089] There are three types of abnormal data for photovoltaic power generation: locally missing values, abnormal fluctuation data, and non-normal data. The methods for handling these three types of bad data are as follows:

[0090] ① Partial missing values: Usually manifested as interruptions in continuous data, with missing data corresponding to some moments. For missing values, dynamic weight correction is used to compensate for the current moment's missing values based on the data of the same weather and time in the previous four days. If a large amount of data is missing in a day, the entire day's data is directly deleted to avoid affecting the final prediction results. The dynamic weight correction method is as follows:

[0091]

[0092] y'1 represents missing values, is the measured value at the same time n days before the missing value. η represents the random weight, and η1+η2+η3+η4=1.

[0093] Abnormal Fluctuation Data: Abnormal fluctuations in PV power generation data fall into two categories: one is caused by sudden changes in weather conditions; the other is caused by abnormal fluctuations in PV power data despite no significant weather fluctuations. The former is normal, while the latter is considered bad data and should be eliminated. When abnormal fluctuations are detected, the weather data is first checked for significant changes. If so, the data is skipped; otherwise, it is eliminated and new data is generated using a dynamic weight correction method.

[0094] ③ For abnormal data of photovoltaic power generation, if the data exceeds the normal range, if it exceeds the maximum power generation power of the photovoltaic power station, it will be treated as abnormal fluctuation data. If a negative number appears, it will be replaced by 0.

[0095] (2) Power load data cleaning

[0096] Bad load data can be divided into three categories: locally missing values, abnormal fluctuation data, and abnormal data. The processing methods are as follows:

[0097] For abnormal load fluctuation data, the density clustering algorithm DBSCAN is used to identify abnormal load fluctuation data in historical data. The identification process is as follows: Figure 2 shown.

[0098] During the DBSCAN algorithm model training process, two important parameters must be specified. One parameter is the minimum sample value, min_samples, which represents the minimum number of sample points required to form a cluster; the other parameter, eps, represents the maximum distance between sample points within the same cluster. If a sample point is not within the clustering circle, measured in eps, it is likely an outlier. For load data cleaning, the minimum sample value, min_samples, is set to 10% of the number of samples in a single identification, and the eps distance is set to 10 times the standard deviation of the sample.

[0099] According to the actual situation of the load history data to be corrected, when there are no abnormal points in the load data of the same type of days on the recent working days or holidays on the day of data cleaning, the daily load of the same type of day is used as the basic curve and the original load data is corrected using the multiple ratio grafting method. Otherwise, the linear interpolation method is used for correction. The formula of the multiple ratio grafting method is as follows:

[0100]

[0101] i=p,p+1,p+2,…,q

[0102] Where x c is the corrected value of the abnormal data point, x t is the value of the base curve, x d is the value of the correct point of the curve to be corrected, and p to q are abnormal data points;

[0103] Linear interpolation mainly uses a straight line passing through (x1, y1) and (x2, y2) to approximate the function The specific formula is as follows.

[0104]

[0105] (2) Distributed photovoltaic power prediction

[0106] Distributed photovoltaic power prediction establishes a model for each distributed photovoltaic station, and performs short-term and ultra-short-term power prediction based on the model. Then, according to the topological relationship and geographical location, the photovoltaic power prediction results at each level of distribution transformer / substation, feeder, substation, district and county, and city are accumulated.

[0107] 1) Short-term power forecast

[0108] The convolutional neural network ResNet18_1D structure is used to build a distributed photovoltaic short-term power prediction model. The photovoltaic historical power, irradiance, temperature and other meteorological data are used as input features, and the predicted daily photovoltaic power is used as the training target to build a short-term power prediction model for distributed photovoltaic short-term power prediction. Based on the grid meteorological data input model provided by the numerical weather forecast, the distributed photovoltaic short-term power prediction results for the next three days are obtained. Based on the input-output model establishment method with limited input delay and the structural characteristics of the convolutional neural network, the one-dimensional convolutional neural network prediction model structure diagram shown in the figure is designed. Figure 3 shown.

[0109] like Figure 3 As shown in the figure, multiple data of a single day are selected to construct a one-dimensional convolutional neural network input-output model for photovoltaic power prediction and extract the time correlation between similar times. 96}, the number of input variables for a single day is 96, Y is the photovoltaic power, and the predicted weather forecast data corresponds to the photovoltaic power of the day.

[0110] The model design and training parameter design are as follows:

[0111] Convolutional layer design: The moving stride is set to 1 and the boundary filling method is selected as SAME.

[0112] Pooling layer design: Select the max pooling method. Taking into account the asymmetry of the input data length and width, use a mixture of 1*2 and 2*2 max pooling kernels with a stride of 2.

[0113] Activation function: Select the LeakyRelu function.

[0114] Fully connected layer design: Keep the dimension consistent with the output Y.

[0115] Network optimization training parameter design: We selected exponentially decayed gradient descent as the gradient descent method, the Adam method as the optimization algorithm, and the SmoothL1Loss loss function using the power predictions and true values. To prevent overfitting, a dropout layer was added to the model.

[0116] Data partitioning: The data partitioning standard is to divide historical data into training data and test data in a ratio of 8:2. All test data are used as verification data for model verification during training, and then input into the network for network training, verification and testing.

[0117] Learning rate decay strategy: using exponential decay method.

[0118] 2) Ultra-short-term power forecast

[0119] A recurrent neural network algorithm is used to train an ultra-short-term irradiance prediction model based on data collected by meteorological monitoring devices. The extreme gradient boosting algorithm (XgBoost) is used to train an ultra-short-term photovoltaic power prediction model. For distributed photovoltaic systems within the meteorological device's range, ultra-short-term irradiance data and daily temperature, humidity, and other data calculated by edge computing at the meteorological monitoring device are input into the ultra-short-term photovoltaic power prediction model for prediction. For distributed photovoltaic systems outside the meteorological device's range, ultra-short-term irradiance prediction data is obtained by combining grid meteorological data with meteorological monitoring device calibration. This data is then input into the ultra-short-term photovoltaic power prediction model for prediction, outputting a distributed photovoltaic ultra-short-term power forecast for the next four hours.

[0120] (3) Industry load characteristics analysis

[0121] The industry load characteristic analysis includes two parts. The first step is to analyze the relationship between the load of each industry and different meteorological factors through historical load data, and to explore the main influencing factors of each industry. The second step is to use the cluster analysis method to extract the typical characteristic curves of similar load days.

[0122] 1) Analysis of influencing factors

[0123] The correlation analysis method is used to obtain the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall, and the main influencing factors of the load of different industries are obtained as a reference for the load forecasting model structure. The correlation coefficient is used as a reference standard for the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall. The calculation formula is shown below:

[0124]

[0125] Where: R is the correlation coefficient; cov(x, y) is the covariance of x and y; D(x) and D(y) are the mean squared errors between the two factors, respectively. The correlation coefficient r ranges from -1 to 1. If r > 0, the two factors are positively correlated; if it is 0, the two factors are negatively correlated. The closer the absolute value of r is to 1, the stronger the linear relationship between factors x and y; the closer the absolute value of r is to 0, the weaker the linear relationship between x and y. Generally, 0 < |r| < 1 indicates a correlation between x and y, but a nonlinear relationship; |r| ≥ 0.7 is considered a high correlation; 0.4 < |r| < 0.7 is considered a significant correlation; and |r| < 0.4 is considered a weak correlation.

[0126] 2) Extract typical characteristic curves

[0127] The k-means clustering algorithm with adaptive k value is used to cluster the historical load of each load for the past year and calculate the typical characteristic curve. The cosine similarity distance is used as the cluster distance indicator. The cosine similarity calculation method is as follows:

[0128]

[0129] Where, P i =[p i1 ,p i2 ,…,p iN ] and P j =[p j1 ,p j2 ,…,p jN ] is the daily load curve of two days i and j, and N is the number of sampling points in one day.

[0130] The silhouette coefficient is selected as the evaluation index of the clustering result. The k value is scanned within a certain range, and the largest silhouette coefficient is selected as the optimal k value. The clustering result at this time is the best. The calculation method of the silhouette coefficient is as follows:

[0131]

[0132] Where: a(i) is the sample P i To other samples P in the same cluster j The average distance of sample P; b(i) is the average distance of sample P i To all samples P of other clusters j M is the number of samples, and the silhouette coefficients of all samples are averaged to obtain the silhouette coefficient S of the overall clustering result.

[0133] (4) Refined load forecasting

[0134] like Figure 4 As shown in the figure, after data cleaning and load characteristics analysis, refined load forecasting includes three steps: ① Establish a load forecasting model for each distribution transformer based on the industry characteristics analysis results; ② Obtain the power load forecast results of the distribution transformer based on the model, and combine them with the photovoltaic power forecast results under the distribution transformer to obtain the refined load forecast results of the distribution transformer; ③ Based on the topological relationship and geographical location, the load forecast results of each level at the city, district, county, substation, and feeder are obtained based on the distribution transformer prediction results. The specific steps are as follows:

[0135] 1) Establish a load forecasting model

[0136] The short-term load forecasting model is built using the recurrent neural network GRU gate unit. The data required for model training include the historical load data of the distribution transformer, the distribution transformer ledger data, including the distribution transformer identification, industry type, capacity, geographical location, topological relationship, etc., as well as historical meteorological data. The model structure is as follows: Figure 5 shown.

[0137] The short-term load forecasting model based on the recurrent neural network GRU gate unit consists of three GRU layers. A Dropout layer is added after each GRU layer to avoid overfitting during model training. The two fully connected layers in the output layer are used to map the features of the GRU gate unit into 96 points of daily load data for the 1×96 predicted output of the output layer.

[0138] The training objective of the model is to predict the daily active power. The input features of the model include:

[0139] a) The historical load curves of the three days before the forecast date and the typical load curves of similar days, with the historical load data normalized;

[0140] b) Valid meteorological data for the three days before the forecast date and similar days, with normalized meteorological data;

[0141] c) Date type: One-hot encoding is used to process date types. The encoding standard is shown in Table 1 below:

[0142] Table 1

[0143]

[0144] 2) Calculate the distribution transformer refined load forecast results

[0145] The historical loads of the three days before the forecast date and the same day last week, as well as the meteorological data and date type of the forecast date, are normalized and one-hot encoded, and then input into the model to obtain the short-term electricity load forecast results for the forecast date and the next two days, a total of three days.

[0146] The ultra-short-term forecast results for the next 0-4 hours are given by correcting the real-time load with the short-term forecast results:

[0147] The sampling time of field data points is generally 15 minutes, so let δt = 15 minutes;

[0148] The ultra-short-term prediction time range is generally 4 hours, so let N = 16;

[0149] The relative error α of short-term load forecasting from the current time t to N δt times before s To predict the relative error α of ultra-short-term prediction us , the ultra-short-term prediction result F at time t+δt us The calculation formula for (t+δt) is as follows:

[0150]

[0151] Where Y(t) and F s (t) represent the actual and short-term forecast values at time t, respectively, W n represents the weight at time tn·δt.

[0152] The electricity load forecast results and the photovoltaic power forecast results are added together to obtain the short-term refined load forecast results of the distribution transformer.

[0153] 3) Calculate the load forecast results at each level

[0154] The short-term refined load forecast results of the distribution transformer are predicted, and the refined load forecast results of the equipment at each level of city, district, county, substation and feeder are accumulated according to the topological relationship.

[0155] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A refined load forecasting method for active distribution networks, characterized in that: The steps include: Step S1, load / power data cleaning: dividing the distribution transformer equivalent load into the power generation component and the power load component of the distributed photovoltaic power data according to the topological structure and the distributed photovoltaic power data; and performing data cleaning on the distributed photovoltaic power data and the power load data respectively; Step S2, distributed photovoltaic power prediction: A model is established for each distributed photovoltaic single station, and short-term and ultra-short-term power predictions are performed based on the model. Then, photovoltaic power prediction results at each level of distribution transformer / substation, feeder, substation, district, county, and city are accumulated based on the topological relationship and geographical location; a distributed photovoltaic short-term power prediction model is built using the convolutional neural network ResNet18_1D structure, and an ultra-short-term irradiance prediction model is trained using a recurrent neural network algorithm based on data collected by meteorological monitoring devices; and an extreme gradient boosting algorithm XgBoost is used to train the photovoltaic power ultra-short-term prediction model. Step S3, industry load characteristics analysis: The industry load characteristics analysis includes two parts. The first step is to analyze the relationship between the load of each industry and different meteorological factors through historical load data, and to explore the main influencing factors of each industry. The second step is to use cluster analysis methods to extract typical characteristic curves of similar load days. Step S4, refined load forecasting: Refined load forecasting includes three steps: ① Establish a load forecasting model for each distribution transformer based on the results of industry characteristic analysis; ② Obtain the power load forecast results of the distribution transformer based on the model, and combine them with the photovoltaic power forecast results under the corresponding distribution transformer to accumulate the refined load forecast results of the corresponding distribution transformer; ③ Based on the topological relationship and geographical location, the load forecast results of each level at the city, district, county, substation, and feeder levels are obtained based on the prediction results of the distribution transformer. The specific implementation steps are as follows: 1) Establish a load forecasting model A short-term load forecasting model is built using the recurrent neural network GRU gate unit; 2) Calculate the distribution transformer refined load forecast results The historical loads of the three days before the forecast date and the same day last week, as well as the meteorological data and date type of the forecast date are normalized and one-hot encoded before being input into the model to obtain the short-term electricity load forecast results for the forecast date and the next two days, a total of three days. The ultra-short-term forecast results for the next 0-4 hours are given by correcting the real-time load with the short-term forecast results: The sampling time of the field data points is 15 minutes, so let δt = 15 minutes; The ultra-short-term prediction time range is 4 hours, and N = 16; The relative error α of short-term load forecasting from the current time t to N δt times before s To predict the relative error α of ultra-short-term prediction us , the ultra-short-term prediction result F at time t+δt us The calculation formula for (t+δt) is as follows: Where Y(t) and F s (t) represent the actual and short-term forecast values at time t, respectively, W n represents the weight at time tn·δt; Add the electricity load forecast results and the photovoltaic power forecast results to obtain the short-term refined load forecast results of the distribution transformer; 3) Calculate the load forecast results at each level The short-term refined load forecast results of the distribution transformer are predicted, and the refined load forecast results of the equipment at each level of city, district, county, substation and feeder are accumulated according to the topological relationship.

2. A refined load forecasting method for active distribution network according to claim 1, characterized in that: The step S1 is specifically implemented as follows: (1) Distributed photovoltaic power data cleaning There are three types of abnormal data for photovoltaic power generation: locally missing values, abnormal fluctuation data, and non-normal data. The methods for handling these three types of bad data are as follows: ① Local missing values: For missing values, dynamic weight correction is used to compensate for the missing values at the current moment based on the data of the same weather and time in the previous four days. If multiple data are missing in one day, the data of the entire day is directly deleted. The dynamic weight correction method is as follows: y'1 represents missing values, is the measured value at the same time in the previous n days with the same weather as the missing value; η i represents random weights, i = 1, 2, 3, 4, and η1 + η2 + η3 + η4 = 1; ② Abnormal fluctuation data: Abnormal fluctuation data of photovoltaic power generation can be divided into two cases: one is the fluctuation of photovoltaic power generation due to sudden changes in weather conditions; the other is the abnormal fluctuation of photovoltaic power data despite no obvious fluctuation in weather conditions. The former is normal, while the latter is bad data and needs to be eliminated. When abnormal fluctuation data is detected, it is necessary to first check whether the meteorological data has changed significantly. If there is a significant change, the data will be skipped. Otherwise, the data will be eliminated and a new data will be corrected using the dynamic weight correction method. ③ For abnormal data, if the data exceeds the normal range and exceeds the maximum power generation of the photovoltaic power station, it will be treated as abnormal fluctuation data. If a negative number appears, it will be replaced by 0; (2) Power load data cleaning Bad load data can be categorized into three types: locally missing values, abnormally fluctuating data, and abnormal data. The processing methods are as follows: For abnormal load fluctuation data, the density clustering algorithm DBSCAN is used to identify abnormal fluctuation data in load history data. During the model training process of the DBSCAN algorithm, there are two important parameters that need to be clarified: one parameter is the minimum sample value min_samples, which represents the minimum number of sample points to form a cluster; the other parameter is the distance eps, which represents the maximum distance between sample points in the same cluster. The minimum sample value min_samples for load data cleaning is set to 10% of the number of samples for a single identification, and the distance eps is set to 10 times the standard deviation of the sample. According to the actual situation of the load history data to be corrected, when there are no abnormal points in the load data of the same type of days on the recent working days or holidays on the day of data cleaning, the daily load of the same type of days is used as the basic curve and the original load data is corrected using the multiple ratio grafting method. Otherwise, the linear interpolation method is used for correction; the formula of the multiple ratio grafting method is as follows: Where x c is the corrected value of the abnormal data point, x t is the value of the base curve, x d is the value of the correct point of the curve to be corrected, and p to q are abnormal data points; Linear interpolation is to use a straight line passing through the points (x1, y1) and (x2, y2) to approximate the function The specific formula is as follows:

3. The method for refined load forecasting for active distribution networks according to claim 1, characterized in that: The step S2 is specifically implemented as follows: 1) Short-term power forecast Using meteorological data including historical photovoltaic power, irradiance, and temperature as input features, and predicted daily photovoltaic power as the training target, a short-term power prediction model is built to perform distributed photovoltaic short-term power prediction; based on the grid meteorological data input model provided by numerical weather forecasting, the distributed photovoltaic short-term power prediction results for the next three days are obtained; based on the input-output model establishment method with limited input delay and the structural characteristics of the convolutional neural network, a one-dimensional convolutional neural network prediction model is designed; multiple data per day are selected to construct a one-dimensional convolutional neural network input-output model for photovoltaic power prediction, and the time correlation between similar times is extracted; where X = {x1, x2, …, x 96 }, the number of input variables for a single day is 96, Y is the photovoltaic power, and the predicted weather forecast data corresponds to the photovoltaic power of the day; The model design and training parameter design are as follows: Convolutional layer design: the moving stride is set to 1, and the boundary filling method is selected as SAME; Pooling layer design: Select the maximum pooling method; taking into account the asymmetry of the length and width of the input data, use a mixture of 1*2 and 2*2 maximum pooling convolution kernels with a stride of 2; Activation function: Select LeakyRelu function; Fully connected layer design: Keep the dimension consistent with the output Y; Network optimization training parameter design: exponential decay gradient descent is selected as the gradient descent method, Adam method is selected as the optimization algorithm, and SmoothL1Loss is calculated using the power prediction value and the actual value as the loss function. At the same time, to prevent overfitting, a dropout layer is added to the model. Data partitioning: The data partitioning standard is to divide historical data into training data and test data in a ratio of 8:

2. All test data is used as validation data for model validation during training, and then input into the network for network training, validation and testing; Learning rate decay strategy: using exponential decay; 2) Ultra-short-term power forecast For distributed photovoltaics within the range of the meteorological device, the ultra-short-term irradiance data and the daily data including temperature and humidity obtained through edge calculation of the meteorological monitoring device are input into the ultra-short-term photovoltaic power prediction model for prediction; for distributed photovoltaics outside the range of the meteorological device, the ultra-short-term irradiance prediction data is obtained through the grid meteorological data + meteorological monitoring device correction method, and then input into the ultra-short-term photovoltaic power prediction model for prediction, and the distributed photovoltaic ultra-short-term power prediction results for the next 4 hours are output.

4. The method for refined load forecasting for active distribution networks according to claim 1, characterized in that: The step S3 is specifically implemented as follows: 1) Analysis of influencing factors The correlation analysis method is used to obtain the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall, and the main influencing factors of the load of different industries are obtained as a reference for the load forecasting model structure; the correlation coefficient is used as a reference standard for the correlation between the load of different industries and temperature, humidity, air pressure, wind speed, wind direction, and rainfall. The calculation formula is shown below: Where: R is the correlation coefficient; cov(x, y) is the covariance of x and y; D(x) and D(y) are the mean square errors between the two factors respectively; the correlation coefficient r value ranges from -1 to 1. If r>0, it means that the two factors are positively correlated; if it is between 0 and 1, it means that the two factors are negatively correlated; the closer the absolute value of r is to 1, the stronger the linear relationship between factors x and y; the closer the absolute value of r is to 0, the weaker the linear relationship between x and y; 0<|r|<1, it means that x and y are correlated, but the relationship is nonlinear; |r|≥0.7 is considered highly correlated; 0.4<|r|<0.7 is considered significantly correlated; |r|<0.4 is considered weakly correlated; 2) Extract typical characteristic curves The k-means clustering algorithm with adaptive k value is used to perform cluster analysis on the historical load of each load in the past year to calculate the typical characteristic curve; the cosine similarity distance is used as the cluster distance indicator, and the cosine similarity calculation method is as follows: Where, P i =[p i1 ,p i2 ,…,p iN ] and P j =[p j1 ,p j2 ,…,p jN ] is the daily load curve of two days i and j, and N is the number of sampling points per day; The silhouette coefficient is selected as the evaluation index of the clustering result. The k value is scanned within the predetermined range, and the largest silhouette coefficient is selected as the optimal k value. The clustering result at this time is optimal. The calculation method of the silhouette coefficient is as follows: Where: a(i) is the sample P i To other samples P in the same cluster j The average distance of sample P; b(i) is the average distance of sample P i To all samples P of other clusters j The average distance of clustering; M is the number of samples, and the silhouette coefficients of all samples are averaged to obtain the silhouette coefficient S of the overall clustering result.

5. The method for refined load forecasting for active distribution network according to claim 1, characterized in that: In step S4 1), The data required for model training includes historical load data of the distribution transformer, distribution transformer inventory data, and historical meteorological data. The short-term load forecasting model based on the recurrent neural network GRU gate unit consists of three GRU layers. A dropout layer is added after each GRU layer to prevent overfitting during model training. The two fully connected layers in the output layer are used to map the features of the GRU gate unit into 96 points of daily load data for the output layer's 1×96 prediction output. The model training objective is to predict the daily active power. The model's input features include: a) The historical load curves of the three days before the forecast date and the typical load curves of similar days, with the historical load data normalized; b) Valid meteorological data for the three days before the forecast date and similar days, with normalized meteorological data; c) Date type: One-hot encoding is used to process date types.

6. A refined load forecasting system for active distribution networks, characterized in that: The method comprises a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, the method steps according to any one of claims 1 to 5 can be implemented.

Images