Feeder line short-term load prediction method and system based on multi-dimensional historical data

Abstract

The invention discloses a feeder line short-term load prediction method and system based on multi-dimensional historical data, and belongs to the field of power systems. The method comprises the following steps: acquiring feeder historical load data, distributed photovoltaic historical power data and historical meteorological data, and preprocessing the data to obtain a historical data set; a load prediction feature construction algorithm is adopted, feature subsets are screened, and multi-dimensional feature data are obtained through mathematical change operation; and inputting the multi-dimensional feature data into the constructed TCN-LSTM model for training to obtain a feeder short-term load prediction model, predicting feeder short-term load according to a set time period in combination with forecast weather, date types and recent historical load data, and giving a prediction result evaluation index to obtain a feeder short-term load prediction value. According to the method, complex factors influencing feeder load changes can be effectively captured, the data quality is improved, the robustness and accuracy of the prediction model are ensured, and the model is made to adapt to the change trend of photovoltaic power generation output and load requirements.

Description

Technical Field

[0001] The present invention belongs to the field of electric power systems, and in particular relates to a method and system for short-term prediction of feeder load based on multi-dimensional historical data. Background Art

[0002] With the transformation of the global energy structure and the rapid development of renewable energy technology, the widespread access of high-proportion distributed power sources and power electronic equipment, the traditional distribution network is undergoing a profound transformation from "passive" to "active". Although this transformation has greatly promoted the diversification and cleanliness of the energy structure, it has also brought unprecedented challenges to the safe, stable and efficient operation of the distribution network. Due to its inherent uncertainty, high volatility and low inertia, distributed new energy has posed a significant threat to the stable operation of the distribution network as its share in the power grid continues to increase.

[0003] As a basic link for safe and stable dispatching of power grids, the accuracy and stability of load forecasting are directly related to the automation level and economic dispatching efficiency of power grid dispatching systems. With the increasingly mature electricity spot market, the accuracy of load forecasting has become a key factor in formulating reliable clearing plans and ensuring stable market operation. At the same time, load forecasting is also an important cornerstone for building a new generation of smart grid dispatching systems, which can provide timely and accurate load demand information for upper-level decision-making and support intelligent and refined dispatching of power grids.

[0004] Faced with the continuous growth of photovoltaic power generation users and the trend of access to distribution networks, the limitations of traditional feeder load forecasting models are gradually emerging. Specifically, given the significant volatility and randomness of photovoltaic power generation output, it will be difficult to accurately capture these dynamic changes by continuing to use traditional forecasting methods, resulting in a significant decrease in forecasting accuracy, making it difficult to meet the strict requirements of modern power systems for load forecasting accuracy.

[0005] Prior art document 1 (CN113902207A) discloses a short-term load forecasting method based on TCN-LSTM. Its shortcomings are that it does not fully consider the impact of distributed photovoltaic power generation on feeder loads, and lacks in-depth mining of multi-dimensional features.

[0006] Prior art document 2 (CN115545319A) discloses a method for short-term load forecasting of power grid based on meteorological similarity day set. Its shortcomings are the lack of in-depth mining and integration of distributed photovoltaic data and historical load data, and the failure to consider the automatic update mechanism of the model. Summary of the invention

[0007] In order to solve the deficiencies in the prior art, the present invention provides a feeder short-term load forecasting method and system based on multi-dimensional historical data. First, historical data required for model training are collected, including feeder load data, photovoltaic power generation data, meteorological data, holiday data, feeder object information, etc.; then, the data is preprocessed, including outlier detection and correction and normalization processing; then, the load forecasting feature construction algorithm is used to screen key features, and multi-dimensional feature data is generated through mathematical transformation; then, the multi-dimensional feature data is input into the TCN-LSTM prediction model for training to build an accurate load forecasting model; finally, according to the forecast weather, date type and recent load history, a periodic rolling prediction mechanism is adopted to generate future load forecast results, and the model performance is verified by detailed evaluation indicators, so as to provide strong support for power grid dispatching and management, and realize accurate prediction of feeder load of distribution system with a high proportion of photovoltaic power generation access.

[0008] The present invention adopts the following technical solutions.

[0009] The first aspect of the present invention discloses a method for short-term feeder load prediction based on multi-dimensional historical data, comprising the following steps:

[0010] Step 1: Obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain a historical data set;

[0011] Step 2: Based on the historical data set in step 1, the load forecast feature construction algorithm is used to screen the feature subset and obtain multi-dimensional feature data through mathematical transformation operations;

[0012] Step 3: Input the multi-dimensional feature data obtained in step 2 into the constructed TCN-LSTM model for training to obtain a feeder short-term load forecasting model;

[0013] Step 4: Based on the feeder short-term load forecasting model obtained in step 3, combined with the forecast weather, date type and recent historical load data, the feeder short-term load is predicted according to the set time period, and the prediction result evaluation index is given to obtain the feeder short-term load forecast value, thus realizing the feeder load short-term forecasting of multi-dimensional historical data.

[0014] Preferably, the specific steps of step 1 include:

[0015] Step 1.1, obtain the historical feeder load and distributed photovoltaic power data on a daily basis to obtain the historical feeder load and distributed photovoltaic power data;

[0016] Step 1.2, obtain historical meteorological information on a daily basis to obtain historical meteorological data;

[0017] Step 1.3, using the subordinate relationship between distributed photovoltaics and feeders, superimposing the distributed photovoltaic power data obtained in step 1.1 on the feeder historical load data obtained in step 1.1 to obtain pure load data, thereby obtaining pure load data, feeder historical load and distributed photovoltaic historical power data;

[0018] Step 1.4, performing horizontal method detection and correction on the abnormal values ​​of the historical meteorological data obtained in step 1.2, the pure load data obtained in step 1.3, the feeder historical load and the distributed photovoltaic historical power data, to obtain the historical data corrected by the horizontal method;

[0019] Step 1.5, obtaining and, based on the date type of each day, performing vertical method detection and correction on the abnormal values ​​of the historical data corrected by the horizontal method obtained in step 1.4, to obtain the historical data corrected by the vertical method;

[0020] Step 1.6, normalize the historical data corrected by the vertical method obtained in step 1.5 to obtain the historical data set.

[0021] Preferably, in step 1.4, if |y(i,t)-y(i,t-1)|>α(t) and |y(i,t)-y(i,t+1)|>β(t), the horizontal method uses the average smoothing calculation method to correct y(i,t), which is expressed as the following formula:

[0022]

[0023] In the formula,

[0024] α(t) and β(t) are thresholds,

[0025] y(i,t) is the value at time t on the i-th day,

[0026] y(i,t) 水 It is the value at time t on the i-th day after correction by the horizontal method.

[0027] Preferably, in step 1.5, if Use the vertical method to correct it, expressed as follows:

[0028]

[0029] Where:

[0030] γ is the threshold,

[0031] y(i,t) 水 is the value at time t on the i-th day after correction by the horizontal method,

[0032] y(i,t) 垂 is the value at time t on the i-th day after correction by the vertical method,

[0033] It is the average value of the historical data at the same time within a set number of days after the horizontal method correction.

[0034] Preferably, the specific steps of step 2 include:

[0035] Step 2.1, solving the change rate of the load at each time interval of the pure load data in the historical data set obtained in step 1, and obtaining the load growth rate;

[0036] Step 2.2, using the correlation coefficient method to solve the correlation coefficient between the historical meteorological data and the distributed photovoltaic power generation data and the pure load data in the historical data set obtained in step 1, and screening out the feeder characteristics according to the set threshold;

[0037] Step 2.3, extracting the feeder characteristic principal components and factors from the feeder characteristic screened in step 2.2, and then combining them with the load growth rate obtained in step 2.1 and the feeder characteristic screened in step 2.2, performing mathematical transformation operations, and obtaining multi-dimensional feeder characteristics.

[0038] Preferably, in step 2.1, the load change rate is expressed by the following formula:

[0039]

[0040] In the formula,

[0041] f (t,t-k) is the rate of change of the load value at time t compared to the load value at time tk,

[0042] y 1,t ,y 1,t-k They are the pure load data values ​​at time t and tk respectively.

[0043] Preferably, in step 2.2, feeder features with a correlation coefficient greater than 0.2 are screened, and the correlation coefficient r xy , expressed as follows:

[0044]

[0045] In the formula,

[0046] n represents the sample size,

[0047] x j is the historical meteorological data and distributed photovoltaic power generation data in the historical data set of the jth sample,

[0048] y j is the pure load data in the historical data set of the jth sample,

[0049] x j andj The mean of .

[0050] Preferably, in step 3, the specific steps include:

[0051] Step 3.1, stack TCN residual blocks to form the backbone of the TCN network, and combine with the dilated convolutional layer to build the TCN network;

[0052] Step 3.2, build the LSTM layer, including the forget gate, input gate, output gate and unit state. After the LSTM layer, add a fully connected layer as the output layer to build the LSTM network.

[0053] Step 3.3, the TCN network constructed in step 3.1 and the LSTM network constructed in step 3.2 are integrated together through a sequential model to obtain a TCN-LSTM model;

[0054] In step 3.4, the multidimensional feature data obtained in step 2 is divided into a training set and a validation set. The TCN-LSTM model obtained in step 3.3 is iteratively trained using the training set. At the end of each training cycle, the performance of the trained TCN-LSTM model on the validation set is evaluated, and the model weights with the best performance are saved to obtain the feeder short-term load forecasting model.

[0055] Preferably, in step 4, the specific steps include:

[0056] Step 4.1, after normalizing the forecast weather, date type and recent historical load data, the data is used as the input of the feeder short-term load forecasting model, and after denormalizing the output value of the feeder short-term load forecasting model, the feeder short-term load forecast value is obtained;

[0057] Step 4.2, performing error analysis on the feeder short-term load forecast value obtained in step 4.1 to obtain a reference error;

[0058] Step 4.3, based on the reference error obtained in step 4.2, calculate and obtain the average daily load forecast accuracy;

[0059] Step 4.4, based on the average daily load forecast accuracy obtained in step 4.3, establish the performance evaluation standard of the feeder short-term load forecasting model within the set time period, obtain the forecast result evaluation index, and realize the feeder load short-term forecasting of multi-dimensional historical data.

[0060] Preferably, in step 4.2, the reference error E of a single object load m in time period p is m,p As shown below:

[0061]

[0062] In the formula,

[0063] The forecast value in period p is the short-term load forecast value of the feeder for a single object load m in period p.

[0064] The actual value in period p is the actual value of the short-term load of the feeder for a single object load m in period p.

[0065] The rated power is the rated power of a single object load m in time period p.

[0066] Preferably, in step 4.3, the average daily load forecast accuracy is expressed by the following formula:

[0067]

[0068] In the formula,

[0069] N is the total number of daily forecast periods,

[0070] A p is the regional statistical error for time period p, expressed as follows:

[0071]

[0072] In the formula,

[0073] M is the total number of loads at all levels that can be assessed in the area,

[0074] E m,p Reference error of a single object load m in time period p.

[0075] Preferably, in step 4.4, a performance evaluation standard of the feeder short-term load forecasting model within a set time period is established, including:

[0076] The average daily load forecast accuracy of each day in the past A days is weighted averaged to obtain the average accuracy, and it is determined whether the average accuracy is less than the set threshold B. If so, the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days are incorporated into the training set, the feeder short-term load forecasting model is retrained, and the model weight is updated. Otherwise, the next determination is performed;

[0077] Count the number of days in the past A days when the average daily load forecast accuracy is lower than the set threshold B, and determine whether the statistical days are greater than the set day threshold C. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecasting model, and update the model weights. Otherwise, proceed to the next determination;

[0078] Sort the average daily load forecast accuracy of the past A days from low to high, select the lowest average daily load forecast accuracy, and determine whether the lowest average daily load forecast accuracy is less than the set threshold D. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecast model, and update the model weights. Otherwise, terminate the incremental training process, and do not train the model in this automatic training cycle.

[0079] A second aspect of the present invention discloses a feeder load short-term prediction system based on multi-dimensional historical data, which is used to run any of the above-mentioned feeder load short-term prediction methods based on multi-dimensional historical data, including:

[0080] Data acquisition module: used to obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain historical data sets;

[0081] Feature extraction module: used to construct an algorithm using load forecasting features, filter feature subsets, and obtain multi-dimensional feature data through mathematical transformation operations;

[0082] Feeder short-term load forecasting model building module: used to input multi-dimensional feature data into the constructed TCN-LSTM model for training and build a feeder short-term load forecasting model;

[0083] Load forecasting module: It is used to forecast the short-term load of feeders in a set time period according to the feeder short-term load forecasting model combined with forecast weather, date type and recent historical load data, and give the forecast result evaluation index to obtain the feeder short-term load forecast value, thus realizing the feeder load short-term forecasting based on multi-dimensional historical data.

[0084] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0085] (1) The present invention integrates feeder load data, photovoltaic power generation output data, weather forecast data, and social behavior pattern data such as holidays, and uses mathematical transformation to generate multidimensional feature data as model input, constructing a comprehensive data analysis framework that effectively captures the complex factors that affect feeder load changes.

[0086] (2) By introducing an abnormal data detection algorithm that combines the horizontal method with the vertical method, the present invention can automatically identify and correct abnormal values ​​in the data set, significantly improve the data quality, ensure the robustness and accuracy of the prediction model, and be able to adapt to the changing trends of photovoltaic power generation output and load demand, providing more powerful support for power grid dispatching and management.

[0087] (3) The present invention adopts the TCN-LSTM deep learning architecture that combines the temporal convolutional network with the long short-term memory network, which enhances the model's ability to learn the local features of time series data, effectively captures long-term dependencies, and improves the model's ability to predict complex dynamic systems.

[0088] (4) The present invention designs a mechanism to regularly expand the training set and update the prediction model, ensuring that the model can continuously learn the latest data features and adapt to the changing trends of photovoltaic power generation output and load demand. The model can be regularly updated according to the accuracy of the prediction results. This adaptive update capability enables the prediction model to always maintain a high prediction accuracy, ensuring the timeliness and accuracy of the prediction model, providing strong support for the dispatching decision of the power system, and improving the load prediction accuracy. At the same time, it also enhances the model's adaptability and prediction ability to grid load changes in an environment with high penetration of distributed energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is the overall flow chart of the present invention;

[0090] Figure 2 It is a flow chart of load forecasting feature construction;

[0091] Figure 3 It is the load prediction model update flow chart. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0093] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0094] refer to Figure 1 As shown, Embodiment 1 of the present invention proposes a feeder short-term load forecasting method based on multi-dimensional historical data, comprising the following steps:

[0095] S1. Obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain a historical load data set.

[0096] The specific process of step S1 includes the following steps:

[0097] Step S11. Obtain the historical feeder load and distributed photovoltaic power generation data at regular intervals every day, with a data time interval of 15 minutes, to obtain the historical feeder load and distributed photovoltaic power generation data;

[0098] In step S11, the database table structure of the feeder historical load and distributed photovoltaic power generation historical data is designed as follows:

[0099] Table 1 Feeder historical load data table

[0100] Serial number Domain Chinese Name Domain English name type 1 date date DATETIME 2 Logo id varchar(22) 3 Update time update_time DATETIME 4 Object Type dtype varchar(22) 5 Measured load 00:00:00 LOAD0 float 6 ... ... ... 7 ... LOAD94 float 8 Measured load 23:45:00 LOAD95 float

[0101] Table 2 Distributed photovoltaic power generation historical data table

[0102] Serial number Domain Chinese Name Domain English name type 1 date date DATETIME 2 Logo id varchar(22) 3 Update time update_time DATETIME 4 Object Type dtype varchar(22) 5 Measured photovoltaic power 00:00:00 LOAD0 float 6 ... ... ... 7 ... LOAD94 float 8 Measured photovoltaic power 23:45:00 LOAD95 float

[0103] Step S12. Obtain gridded historical meteorological information on a daily basis, including meteorological elements such as temperature, humidity, rainfall, irradiance, air pressure, wind speed, etc. The data interval is 15 minutes, and historical meteorological data is obtained;

[0104] In step S12, the database table structure of historical meteorological data is designed as follows:

[0105] Table 3 Historical meteorological data

[0106] Serial number Domain Chinese Name Domain English name type 1 Logo id varchar(50) 2 time time timestamp(6) 3 Forecast time fbtime timestamp(6) 4 Radiation downward from the surface swdown dec(11,3) 5 Direct downward radiation from the surface swddir dec(11,3) 6 Diffuse Radiation swddif dec(11,3) 7 Total cloud cover totalcloud dec(11,3) 8 Low-level cloud cover lowcloud dec(11,3) 9 temperature temp dec(11,3) 10 Relative humidity humid dec(11,3) 11 10M wind speed sp10 dec(11,3) 12 10M Wind Direction dir10 dec(11,3) 13 pressure pres dec(11,3) 14 Rainfall rainfall dec(11,3) 15 longitude lon dec(11,3) 16 latitude lat dec(11,3)

[0107] Step S13. Using the subordinate relationship between distributed photovoltaics and feeders, the distributed photovoltaic power data obtained in step S11 is superimposed on the feeder historical load data obtained in step S11 to obtain pure load data, thereby obtaining pure load data, feeder historical load and distributed photovoltaic historical power data.

[0108] In step S13, the calculation formula of pure load data is as follows:

[0109] Pure load data = historical feeder load data + distributed photovoltaic power (1)

[0110] For example: the historical load value at 12 o'clock today is 1, and the distributed photovoltaic power value is 0.5, then the pure load data at 12 o'clock today is 1.5.

[0111] Step S14. Perform horizontal method detection and correction on the abnormal values ​​of the historical meteorological data obtained in step S12 and the pure load data, feeder historical load and distributed photovoltaic historical power data obtained in step S13 to obtain historical data.

[0112] In step S14, the load is continuous. Under normal circumstances, the load is a continuous and stable sequence without sudden changes. Therefore, the load at a certain moment will not differ too much from the loads before and after. If the load number variable at a certain moment exceeds the threshold range, the average value method can be used for smoothing calculation.

[0113] If |y(i, t)-y(i, t-1)|>α(t) and |y(i, t)-y(i, t+1)|>β(t), the horizontal method uses the average smoothing method to correct y(i, t), which is expressed as follows:

[0114]

[0115] In the formula,

[0116] α(t) and β(t) are thresholds,

[0117] y(i, t) is the value at time t on the i-th day,

[0118] y(i,t) 水 It is the value at time t on the i-th day after correction by the horizontal method.

[0119] Step S15. Obtain and perform vertical method detection and correction on the abnormal values ​​of the historical data obtained in S14 according to the date type of each day to obtain the historical data processed by the vertical method.

[0120] Daily date types include three types: weekdays, weekends, and holidays.

[0121] In step S15, the load has a strong periodicity, so according to the period, it can be known that the load values ​​of the same date type at the same time should be relatively similar. When the difference in load values ​​between the two exceeds the threshold, the vertical processing method is used to correct the abnormal data that exceeds the range.

[0122] like Use the vertical method to correct it, expressed as follows:

[0123]

[0124] Where:

[0125] γ is the threshold,

[0126] y(i,t) 水 is the value at time t on the i-th day after correction by the horizontal method,

[0127] y(i,t) 垂 is the value at time t on the i-th day after correction by the vertical method,

[0128] It is the average value of the historical data at the same time within the set number of days in the most recent few days after the horizontal method correction. In practice, the most recent few days are generally 3-7 days.

[0129] Step S16: normalize the historical data processed by the vertical method obtained in step S15 to obtain a historical data set.

[0130] In step S16, normalization processing can not only reduce the range of sample data, but also reduce the training time of the model. The model is trained with normalized data, which can not only accelerate the convergence speed of the model, but also further improve the prediction accuracy of the model. The historical data processed by the vertical method is normalized, and the specific calculation formula is as follows:

[0131]

[0132] Where:

[0133] Represents normalized historical data, with a size range of [0, 1].

[0134] x represents the historical data before normalization,

[0135] x max and x min They respectively represent the maximum and minimum values ​​in the historical load data after being processed by the vertical method with the same dimension.

[0136] S2. Based on the historical data set obtained in S1, a load forecasting feature construction algorithm is used to screen useful feature subsets and form multi-dimensional feature data through mathematical transformation operations.

[0137] The specific process of step S2 includes the following steps:

[0138] Step S21. Extract features from the pure load data in the historical data set, construct different time series features, and obtain the load growth rate.

[0139] In step S21, the load change rate is defined, which is characterized by the change rate between loads at different times in the load sequence, that is,

[0140]

[0141] In the formula,

[0142] f (t,t-k) is the growth rate of the load value at time t compared to the load value at time tk;

[0143] y 1,t ,y 1,t-k They are the pure load data values ​​at time t and tk respectively.

[0144] Step S22. Use the correlation coefficient method to study the impact of multiple variables on load data, solve the correlation coefficient between the historical meteorological data and distributed photovoltaic power generation data and the pure load data in the historical data set obtained in step 1, and screen out the feeder characteristics based on the set threshold.

[0145] In step S22, based on the processed input data, a load forecasting feature construction algorithm is used to screen useful feature subsets by analyzing the correlation between the features and the feeder load. The correlation coefficient method is used to study the impact of multiple variables on load data. Feature selection is based on data features such as meteorology and distributed photovoltaic power, combined with methods such as the correlation coefficient method to achieve the selection of feature subsets useful for load forecasting. The correlation coefficient method is a filtering method in feature selection. By calculating the correlation coefficient between each feature and the predicted load, the correlation between each feature and the load is evaluated, and features that are highly correlated with the target variable are selected as input variables of the model. The correlation coefficient method is used to study the impact of multiple variables on load data and determine the intrinsic connection between variables and load changes. The Pearson correlation coefficient method correlation coefficient r xy The formula is:

[0146]

[0147] In the formula,

[0148] n represents the sample size,

[0149] x j is the historical meteorological data and distributed photovoltaic power generation data in the historical data set of the jth sample;

[0150] y j is the pure load data in the historical data set of the jth sample;

[0151] x j and j The mean of .

[0152] The degree of correlation of variables is determined by the absolute value of the correlation coefficient. The larger the absolute value, the stronger the correlation. Therefore, feature variables with a correlation coefficient greater than 0.2 are selected for model training.

[0153] Step S23. Extract the feeder feature principal components and factors from the feeder features screened in S22, derive new features, and then combine them with the load growth rate obtained in S21 and the feeder features screened in S22, perform mathematical transformation operations, and obtain multi-dimensional feeder features.

[0154] In step S23, new features are derived from the feature subsets filtered by S22. For all feature values, i.e., the feature subsets filtered by S22, the new features derived from the feature subsets filtered by S22, and the load growth rate obtained by S21, mathematical transformation operations such as logarithm, square, inverse, one-hot encoding, etc. are used to form multi-dimensional feature data. Feeder load forecast feature construction process Figure 2 shown.

[0155] S3. Input the multi-dimensional feature data in S2 into the TCN-LSTM model for training and build a feeder short-term load forecasting model.

[0156] The specific process of step S3 includes the following steps:

[0157] Step S31. Design a TCN residual block, stack multiple TCN residual blocks to form the backbone of the TCN network, and combine them with the dilated convolutional layer to obtain the TCN network.

[0158] In step S31, the TCN residual block is the core component of TCN, and multiple TCN residual blocks are stacked to form the backbone of the TCN network. Each residual block contains two causal convolutional layers with an expansion convolutional layer between them. This design allows TCN to capture dependencies in long sequences. The expansion convolution can be expressed as:

[0159]

[0160] in,

[0161] y t is the output of the convolution operation at time t,

[0162] w k is the weight of the convolution kernel when the convolution kernel is k,

[0163] x t-d·k is the element of the input sequence at time td·k considering the dilation rate d,

[0164] d is the expansion rate,

[0165] K is the size of the convolution kernel.

[0166] Step S32. Design an LSTM layer, including a forget gate, an input gate, an output gate, and a unit state. After the LSTM layer, add a fully connected layer as an output layer to obtain an LSTM network.

[0167] In step S32, the LSTM layer is the core of the long short-term memory network, which effectively captures long-term dependencies by introducing a forget gate, an input gate, an output gate, and a unit state.

[0168]

[0169] In the formula,

[0170] σ and tanh are Sigmoid and hyperbolic tangent functions respectively,

[0171] W fx , W ix , W ox With b f , b i , b o are the corresponding weight coefficient matrix and bias passing through the LSTM layer, i.e., the parameters adjusted after feedback from the loss function during training;

[0172] f t Forget gating, which controls the S of the previous state t-1 The specific forgotten part;

[0173] i t is the input gate, which controls the input x t Memory;

[0174] o t It is output gating, which controls the output of the current state.

[0175] f t 、i t , o t It is converted into a value between 0 and 1 by multiplying the splicing vector by the above structural parameter matrix through the sigmoid activation function;

[0176] S t The result is converted to a value between –1 and 1 through the nonlinear function tanh;

[0177] h t is the output of step t.

[0178] Step S33. Integrate the TCN network obtained in step S31 and the LSTM network obtained in step S32 through a sequential model to form a complete TCN-LSTM model.

[0179] In step S33, the output of TCN is used as the input of the LSTM layer, a dropout layer is added between TCN and LSTM, and a fully connected layer is used after the LSTM layer to give the prediction result. The number of TCN layers is set to 2, the number of convolution kernels is set to 64, the convolution kernel size is set to 2, the expansion coefficient is set to (1, 2, 4), and the Dropout1 is set to 0.5; the number of LSTM layers is set to 3, each layer contains 512, 256, and 128 neurons respectively; the Dense layer is set to 1 layer; the model activation function uses the Relu function, and the loss function uses the MAE.

[0180] Step S34. Use the training set data to iteratively train the TCN-LSTM network. At the end of each training cycle, evaluate the performance of the model on the validation set, and save the model weights with the best performance to obtain the feeder short-term load forecasting model.

[0181] The multidimensional feature data is divided into a training set and a validation set in a ratio of 80% and 20%.

[0182] In step S34, at the end of each training cycle, the performance of the model on the validation set is evaluated, and the model weights of the best performance are saved. In this way, the performance changes of the model during the training process can be monitored, and when the model reaches the best performance, its weights can be saved for subsequent testing and application.

[0183] S4. Based on the TCN-LSTM model constructed in S3, combined with forecast meteorological data, date type, recent historical load and other data, the future load forecast results are predicted on a rolling basis, and the evaluation indicators of the model prediction results are given.

[0184] The specific process of step S4 includes the following steps:

[0185] Step S41: Use the trained model to obtain the feeder short-term load forecast value.

[0186] In step S41, the forecast meteorological data, date type, recent historical load and other data are normalized and used as the input of the TCN-LSTM model. The output value of the TCN-LSTM model is denormalized to obtain the feeder short-term load forecast value.

[0187] Step S42: Perform error analysis on the feeder short-term load forecast value obtained in S41 to obtain a reference error.

[0188] In step S42, the reference error E of a single object load m in time period p is m,p As shown below:

[0189]

[0190] In the formula,

[0191] The forecast value in period p is the short-term load forecast value of the feeder for a single object load m in period p.

[0192] The actual value in period p is the actual value of the short-term load of the feeder for a single object load m in period p.

[0193] The rated power is the rated power of a single object load m in time period p.

[0194] Step S43: Calculate and obtain the average daily load forecast accuracy based on the reference error obtained in S42.

[0195] In step S43, the average daily load prediction accuracy is expressed by the following formula:

[0196]

[0197] In the formula,

[0198] N is the total number of daily forecast periods,

[0199] A p is the regional statistical error for time period p, expressed as follows:

[0200]

[0201] In the formula,

[0202] M is the total number of loads at all levels that can be assessed in the area,

[0203] E m,p Reference error of a single object load m in time period p.

[0204] Step S44. According to the average daily load forecast accuracy obtained in S43, the performance evaluation standard of the feeder short-term load forecast model in the set time period is established, and the forecast result evaluation index is obtained to realize the feeder short-term load forecast based on multi-dimensional historical data. In step S44, the performance evaluation standard of the model in a specific forecast time period is established to determine the model update method. When the model automatic training time period is reached, the incremental training process is entered.

[0205] The average daily load forecast accuracy of each day in the past A days is weighted averaged to obtain the average accuracy, and it is determined whether the average accuracy is less than the set threshold B. If so, the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days are incorporated into the training set, the feeder short-term load forecasting model is retrained, and the model weight is updated. Otherwise, the next determination is performed;

[0206] Count the number of days in the past A days when the average daily load forecast accuracy is lower than the set threshold B, and determine whether the statistical days are greater than the set day threshold C. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecasting model, and update the model weights. Otherwise, proceed to the next determination;

[0207] Sort the average daily load forecast accuracy of the past A days from low to high, select the lowest average daily load forecast accuracy, and determine whether the lowest average daily load forecast accuracy is less than the set threshold D. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecast model, and update the model weights. Otherwise, terminate the incremental training process, and do not train the model in this automatic training cycle.

[0208] Process such as Figure 3 As shown, the specific process is as follows:

[0209] (1) First, calculate the average accuracy rate in the past 30 days, that is, take the weighted average of the average daily load forecast accuracy rate and determine whether it is less than 90%. If so, directly add new data to the training set and retrain the model.

[0210] Update the model weights, otherwise proceed to the next decision.

[0211] (2) Count the number of days in the past 30 days when the average daily load forecast accuracy is less than 90%, and determine whether the number of statistical days is greater than 5. If so, directly add new data to the training set, retrain the model, and update the model weights. Otherwise, proceed to the next determination.

[0212] (3) Sort the average daily load forecast accuracy of the past 30 days from low to high, select the lowest average daily load forecast accuracy, and determine whether the lowest average daily load forecast accuracy is less than 70%. If so, directly add new data to the training set, retrain the model, and update the model weights. Otherwise, terminate the incremental training process and do not train the model in this automatic training cycle.

[0213] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0214] (1) The present invention integrates feeder load data, photovoltaic power generation output data, weather forecast data, and social behavior pattern data such as holidays, and uses mathematical transformation to generate multidimensional feature data as model input, constructing a comprehensive data analysis framework that effectively captures the complex factors that affect feeder load changes.

[0215] (2) By introducing an abnormal data detection algorithm that combines the horizontal method with the vertical method, the present invention can automatically identify and correct abnormal values ​​in the data set, significantly improve the data quality, ensure the robustness and accuracy of the prediction model, and be able to adapt to the changing trends of photovoltaic power generation output and load demand, providing more powerful support for power grid dispatching and management.

[0216] (3) The present invention adopts the TCN-LSTM deep learning architecture that combines the temporal convolutional network with the long short-term memory network, which enhances the model's ability to learn the local features of time series data, effectively captures long-term dependencies, and improves the model's ability to predict complex dynamic systems.

[0217] (4) The present invention designs a mechanism to regularly expand the training set and update the prediction model, ensuring that the model can continuously learn the latest data features and adapt to the changing trends of photovoltaic power generation output and load demand. The model can be regularly updated according to the accuracy of the prediction results. This adaptive update capability enables the prediction model to always maintain a high prediction accuracy, ensuring the timeliness and accuracy of the prediction model, providing strong support for the dispatching decision of the power system, and improving the load prediction accuracy. At the same time, it also enhances the model's adaptability and prediction ability to grid load changes in an environment with high penetration of distributed energy.

[0218] Embodiment 2 of the present invention proposes a feeder short-term load forecasting system based on multi-dimensional historical data, which is used to run a feeder short-term load forecasting method based on multi-dimensional historical data proposed in Embodiment 1, including:

[0219] Data acquisition module: used to obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain historical data sets;

[0220] Feature extraction module: used to construct an algorithm using load forecasting features, filter feature subsets, and obtain multi-dimensional feature data through mathematical transformation operations;

[0221] Feeder short-term load forecasting model building module: used to input multi-dimensional feature data into the constructed TCN-LSTM model for training and build a feeder short-term load forecasting model;

[0222] Load forecasting module: It is used to forecast the short-term load of feeders in a set time period according to the feeder short-term load forecasting model combined with forecast weather, date type and recent historical load data, and give evaluation indicators for the forecast results to obtain the feeder short-term load forecast value, thus realizing the short-term forecast of feeder load based on multi-dimensional historical data.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for short-term feeder load prediction based on multi-dimensional historical data, characterized in that: The following steps are involved: Step 1: Obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain a historical data set; Step 2: Based on the historical data set in step 1, the load forecast feature construction algorithm is used to screen the feature subset and obtain multi-dimensional feature data through mathematical transformation operations; Step 3: Input the multi-dimensional feature data obtained in step 2 into the constructed TCN-LSTM model for training to obtain a feeder short-term load forecasting model; Step 4: Based on the feeder short-term load forecasting model obtained in step 3, combined with the forecast weather, date type and recent historical load data, the feeder short-term load is predicted according to the set time period, and the prediction result evaluation index is given to obtain the feeder short-term load forecast value, thus realizing the feeder load short-term forecasting of multi-dimensional historical data.

2. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 1 is characterized in that: The specific steps of step 1 include: Step 1.1, obtain the historical feeder load and distributed photovoltaic power data on a daily basis to obtain the historical feeder load and distributed photovoltaic power data; Step 1.2, obtain historical meteorological information on a daily basis to obtain historical meteorological data; Step 1.3, using the subordinate relationship between distributed photovoltaics and feeders, superimposing the distributed photovoltaic power data obtained in step 1.1 on the feeder historical load data obtained in step 1.1 to obtain pure load data, thereby obtaining pure load data, feeder historical load and distributed photovoltaic historical power data; Step 1.4, performing horizontal method detection and correction on the abnormal values ​​of the historical meteorological data obtained in step 1.2, the pure load data obtained in step 1.3, the feeder historical load and the distributed photovoltaic historical power data, to obtain the historical data corrected by the horizontal method; Step 1.5, obtaining and, based on the date type of each day, performing vertical method detection and correction on the abnormal values ​​of the historical data corrected by the horizontal method obtained in step 1.4, to obtain the historical data corrected by the vertical method; Step 1.6, normalize the historical data corrected by the vertical method obtained in step 1.5 to obtain the historical data set.

3. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 2 is characterized in that: In step 1.4, if |y(i,t)-y(i,t-1)|>α(t) and |y(i,t)-y(i,t+1)|>β(t), the horizontal method uses the average smoothing method to correct y(i,t), which is expressed as the following formula: In the formula, α(t) and β(t) are thresholds, y(i,t) is the value at time t on the i-th day, y(i,t) 水 It is the value at time t on the i-th day after correction by the horizontal method.

4. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 2 is characterized in that: In step 1.5, if Use the vertical method to correct it, expressed as follows: Where: γ is the threshold, y(i,t) 水 is the value at time t on the i-th day after correction by the horizontal method, y(i,t) 垂 is the value at time t on the i-th day after correction by the vertical method, It is the average value of the historical data at the same time within a set number of days after the horizontal method correction.

5. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 1 is characterized in that: The specific steps of step 2 include: Step 2.1, solving the change rate of the load at each time interval of the pure load data in the historical data set obtained in step 1, and obtaining the load growth rate; Step 2.2, using the correlation coefficient method to solve the correlation coefficient between the historical meteorological data and the distributed photovoltaic power generation data and the pure load data in the historical data set obtained in step 1, and screening out the feeder characteristics according to the set threshold; Step 2.3, extracting the feeder characteristic principal components and factors from the feeder characteristic screened in step 2.2, and then combining them with the load growth rate obtained in step 2.1 and the feeder characteristic screened in step 2.2, performing mathematical transformation operations, and obtaining multi-dimensional feeder characteristics.

6. The method for short-term feeder load forecasting based on multi-dimensional historical data according to claim 5 is characterized in that: In step 2.1, the load change rate is expressed as follows: In the formula, f (t,t-k) is the rate of change of the load value at time t compared to the load value at time tk, y 1,t ,y 1,t-k They are the pure load data values ​​at time t and tk respectively.

7. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 5 is characterized in that: In step 2.2, the feeder features with correlation coefficient greater than 0.2 are screened. The correlation coefficient r xy , expressed as follows: In the formula, n represents the sample size, x j is the historical meteorological data and distributed photovoltaic power generation data in the historical data set of the jth sample, y j is the pure load data in the historical data set of the jth sample, x j and j The mean of .

8. The method for short-term feeder load forecasting based on multi-dimensional historical data according to claim 1 is characterized in that: In step 3, the specific steps include: Step 3.1, stack TCN residual blocks to form the backbone of the TCN network, and combine with the dilated convolutional layer to build the TCN network; Step 3.2, build the LSTM layer, including the forget gate, input gate, output gate and unit state. After the LSTM layer, add a fully connected layer as the output layer to build the LSTM network. Step 3.3, the TCN network constructed in step 3.1 and the LSTM network constructed in step 3.2 are integrated together through a sequential model to obtain a TCN-LSTM model; In step 3.4, the multidimensional feature data obtained in step 2 is divided into a training set and a validation set. The TCN-LSTM model obtained in step 3.3 is iteratively trained using the training set. At the end of each training cycle, the performance of the trained TCN-LSTM model on the validation set is evaluated, and the model weights with the best performance are saved to obtain the feeder short-term load forecasting model.

9. The method for short-term feeder load forecasting based on multi-dimensional historical data according to claim 1 is characterized in that: In step 4, the specific steps include: Step 4.1, after normalizing the forecast weather, date type and recent historical load data, the data is used as the input of the feeder short-term load forecasting model, and after denormalizing the output value of the feeder short-term load forecasting model, the feeder short-term load forecast value is obtained; Step 4.2, performing error analysis on the feeder short-term load forecast value obtained in step 4.1 to obtain a reference error; Step 4.3, based on the reference error obtained in step 4.2, calculate and obtain the average daily load forecast accuracy; Step 4.4, based on the average daily load forecast accuracy obtained in step 4.3, establish the performance evaluation standard of the feeder short-term load forecasting model within the set time period, obtain the forecast result evaluation index, and realize the feeder load short-term forecasting of multi-dimensional historical data.

10. The method for short-term feeder load forecasting based on multi-dimensional historical data according to claim 9, characterized in that: In step 4.2, the reference error E of a single object load m in time period p is m,p As shown below: In the formula, The forecast value in period p is the short-term feeder load forecast value of a single object load m in period p. The actual value in period p is the actual value of the short-term load of the feeder for a single object load m in period p. The rated power is the rated power of a single object load m in time period p.

11. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 9, characterized in that: In step 4.3, the average daily load forecast accuracy is expressed as follows: In the formula, N is the total number of daily forecast periods, A p is the regional statistical error for time period p, expressed as follows: In the formula, M is the total number of loads at all levels that can be assessed in the area, E m,p Reference error of a single object load m in time period p.

12. The method for short-term feeder load prediction based on multi-dimensional historical data according to claim 9, characterized in that: In step 4.4, the performance evaluation criteria of the feeder short-term load forecasting model within a set time period are established, including: The average daily load forecast accuracy of each day in the past A days is weighted averaged to obtain the average accuracy, and it is determined whether the average accuracy is less than the set threshold B. If so, the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days are incorporated into the training set, the feeder short-term load forecasting model is retrained, and the model weight is updated. Otherwise, the next determination is performed; Count the number of days in the past A days when the average daily load forecast accuracy is lower than the set threshold B, and determine whether the statistical days are greater than the set day threshold C. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecasting model, and update the model weights. Otherwise, proceed to the next determination; Sort the average daily load forecast accuracy of the past A days from low to high, select the lowest average daily load forecast accuracy, and determine whether the lowest average daily load forecast accuracy is less than the set threshold D. If so, incorporate the historical load data, distributed photovoltaic power data, historical meteorological data and multi-dimensional feature data of the past A days into the training set, retrain the feeder short-term load forecast model, and update the model weights. Otherwise, terminate the incremental training process, and do not train the model in this automatic training cycle.

13. A feeder load short-term prediction system based on multi-dimensional historical data, used for running a feeder load short-term prediction method based on multi-dimensional historical data according to any one of claims 1 to 9, characterized in that: Data acquisition module: used to obtain historical feeder load data, distributed photovoltaic historical power data and historical meteorological data, pre-process the data, and obtain historical data sets; Feature extraction module: used to construct an algorithm using load forecasting features, filter feature subsets, and obtain multi-dimensional feature data through mathematical transformation operations; Feeder short-term load forecasting model building module: used to input multi-dimensional feature data into the constructed TCN-LSTM model for training and build a feeder short-term load forecasting model; Load forecasting module: It is used to forecast the short-term load of feeders in a set time period according to the feeder short-term load forecasting model combined with forecast weather, date type and recent historical load data, and give evaluation indicators for the forecast results to obtain the feeder short-term load forecast value, thus realizing the short-term forecast of feeder load based on multi-dimensional historical data.

Images