CN117273802B - Medium-short term electricity sales quantity prediction method based on deep neural network - Google Patents

Abstract

The invention discloses a method for predicting middle-short-term electricity sales based on a deep neural network, which relates to the technical field of electricity sales data processing and comprises the following steps: acquiring an original sample set of each area; performing data cleaning on the original sample set; preprocessing the cleaned data to obtain a preprocessed sample set; the management center carries out prediction optimization coefficient YH analysis on the pretreatment sample sets of each platform region to obtain a prediction optimization sequence of the pretreatment sample sets, and data processing efficiency is improved; after receiving the pretreatment sample set, the prediction terminal processes the time sequence in a sequence difference mode, and an LSTM neural network model is established; initializing the weight of a neural network and the parameters of a particle swarm optimization algorithm by adopting an Nguyen-Widry method; performing model training on the neural network model, and performing model evaluation through a loss function to obtain an optimal medium-short-term electricity selling data prediction model with minimum overall error of a training sample; and the data prediction accuracy is improved.

Description

A short- and medium-term electricity sales forecast method based on deep neural network

Technical field

The invention relates to the technical field of electricity sales data processing, specifically a short- and medium-term electricity sales prediction method based on a deep neural network.

Background technique

With the development of the power market and the increase in user needs, the safe and economical operation of the power grid has become crucial. Accurate short-term forecasting of electricity sales in Taiwan can effectively ensure the safe operation of the power grid, reduce power generation costs, meet user needs and improve social and economic benefits. Since electricity sales in Taiwan have obvious cyclical characteristics and the influencing factors are complex, such as user electricity consumption behavior, load changes, seasonal changes, holidays, etc., it is necessary to choose advanced and accurate short- and medium-term electricity sales forecasting methods.

In the process of power sales forecasting, it is more common to use two neural network models: Recurrent Neural Network (RNN) or Long Short-term Memory Network (LSTM). Given sufficient time, RNN/LSTM can meet most power requirements. However, in the network structure of RNN/LSTM, the input of the current layer is the output of the previous layer. All RNN/LSTM are more suitable for processing time series problems, but it is precisely because of this serial structure that limits RNN/LSTM. The training speed of the model cannot be parallelized, which seriously affects the modeling efficiency when the amount of data is large and the construction period is tight. However, the support vector method based on statistical theory has to determine its optional parameters and functions. The need to rely on manual experience will also affect the prediction effect. Based on the above shortcomings, the present invention proposes a short- and medium-term power sales prediction method based on deep neural networks.

Contents of the invention

The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a short- and medium-term power sales prediction method based on deep neural networks.

In order to achieve the above objectives, according to an embodiment of the first aspect of the present invention, a short- and medium-term power sales prediction method based on a deep neural network is proposed, which includes the following steps:

Step 1: Obtain the original sample set of each station; perform data cleaning on the original sample set; preprocess the cleaned data to obtain the preprocessed sample set and cache it to the management center; the original sample set includes date data, Climate text data and electricity sales historical data;

Step 2: The management center analyzes the prediction optimization coefficient YH of the preprocessing sample sets in each station area, and sorts the preprocessing sample sets according to the size of the prediction optimization coefficient YH to obtain the prediction optimization sequence of the preprocessing sample set;

Then, according to the prediction optimization sequence, the preprocessed sample sets of each station area are sent to the prediction terminal for electricity sales data prediction model training;

Step 3: After the prediction terminal receives the preprocessed sample set, it splits the preprocessed sample set into mutually exclusive training sets, verification sets, and test sets; uses sequence difference to process the time series and establishes an LSTM neural network model; where, Specify the number of input nodes of the long short-term memory neural network LSTM according to the number of input variables, set the appropriate number of hidden layer nodes, and the number of output nodes representing the predicted value of electricity sales;

Step 4: Initialize the neural network weights and particle swarm optimization algorithm parameters; wherein, the Nguyen-Widrow method is used to initialize the neural network weights; perform model training on the neural network model to obtain a trained short- and medium-term electricity sales data prediction model;

Step 5: After obtaining the short- and medium-term electricity sales data prediction model with optimal parameters, use the electricity sales data in the time zone before the time zone to be predicted and the climate text data in the time zone to be predicted as input to the model to obtain the predicted value of electricity sales in the time zone to be predicted.

Further, the management center conducts prediction optimization coefficient YH analysis on the preprocessed sample sets of each station area. The specific analysis steps are:

Obtain the station area corresponding to the preprocessed sample set and obtain the power supply area of the station area; count the length of the power supply line in the power supply area as L1, the number of power supply households as HL, and the average power consumption per household as DL;

The regional correlation coefficient GD of the station area is calculated using the formula GD=L1×g2+HL×g3+DL×g4, where g2, g3, and g4 are preset coefficient factors;

Within a preset time period, the supply and sales error data of each sampling interval of the station area is collected; the supply and sales error data is the difference data between the power supply and the electricity sales; the difference data is taken as a positive number;

Mark the supply and sale error data of each sampling interval as Wi; compare the supply and sale error data Wi with the preset error threshold; if Wi > the preset error threshold, it indicates that the corresponding station area has a large power loss, which increases the power generation. cost, causing additional losses to the operation of the power grid;

The proportion of statistical supply and sale error data Wi > the preset error threshold is Zb. When Wi > the preset error threshold, the difference between Wi and the preset error threshold is obtained and summed to obtain the total excess error value LZ; use the formula GS=Zb×g1+LZ×g5 is calculated to obtain the power supply loss coefficient GS of the station area, where g1 and g5 are both preset coefficient factors;

The regional correlation coefficient and power supply loss coefficient are normalized and their values are calculated, and the prediction optimization coefficient YH of the preprocessed sample set is calculated using the formula YH=ƒ×(GD×b1+GS×b2), where b1, b2 are all preset coefficient factors; ƒ is the preset equalization coefficient.

Further, data cleaning is performed on the original sample set, including: filling or discarding empty values; deduplicating duplicate data; cleaning climate text data with wrong ranges; and performing data verification on climate text data.

Further, preprocess the cleaned data, including:

Perform unit conversion processing on the historical electricity sales data, complete the sampling time points to ensure their continuity, and use the average interpolation method to fill in the missing data at the sampling points to obtain the Taiwan area electricity sales time series; among them, if the sampling time points and data are missing in large areas, Data from other years during the same period are used for filling.

Further, the time series is processed using sequence difference. The specific steps include:

Based on the monthly and quarterly data of the time series, the 12-month or 4-quarter moving average of electricity sales in Taiwan is obtained, and the long-term moving average trend data Q is obtained;

According to the multiplicative model, calculate Y/Q=S×C×I; where Y represents the year, S represents the seasonal component, C represents the periodic component, and I represents the irregular component;

Recalculate Y/Q based on the same month or quarter of each year and take the average to obtain the arithmetic mean Pi of the same month or quarter of each year;

Taking the arithmetic mean Pi of the same month or quarter in each year as the numerator and the sum of the arithmetic means of all months or quarters as the denominator, the seasonal ratio Si of each month or quarter is calculated, where Si=(N×Pi )/(P1+P2+…+PN); seasonal ratio Si is the correction coefficient of seasonal factors to the long-term trend of electricity sales; P1+P2+…+PN is the sum of the arithmetic average of all months or quarters; N is the number of samples ;

The product of Tt with seasonal factors removed and the corresponding seasonal Si in period t is calculated, which is the predicted value of electricity sales corresponding to period t; where Qt is the moving average trend data of period t.

Further, the neural network weight includes at least one of the input value of the model, hidden value, number of model layers, neuron dropout rate, activation function, initial value of feature weight, and bias.

Further, model training is performed on the neural network model to obtain a trained medium- and short-term electricity sales data prediction model, which specifically includes:

The training set, verification set, and test set are input into the LSTM neural network model as historical feature values for model training, and the model is evaluated through the loss function to obtain the optimal short- and medium-term electricity sales data that minimizes the overall error of the training sample. Predictive model.

Further, model evaluation is performed through the loss function, including:

The root mean square error equation is used as the loss function, and the R2_score coefficient of determination is used as the model fitting evaluation method to evaluate the loss between the model prediction value and the true value;

Determine the trained short- and medium-term electricity sales data prediction model based on the loss assessment results.

Compared with the prior art, the beneficial effects of the present invention are:

1. The present invention obtains the original sample set of each station area; performs data cleaning on the original sample set; preprocesses the cleaned data to obtain the preprocessed sample set and caches it to the management center; the management center preprocesses each station area The sample set is analyzed for the prediction optimization coefficient YH, and the prediction optimization coefficient YH of the preprocessed sample set is calculated based on the length of the power supply line in the power supply area, the number of households with power supply, average electricity consumption per household, and the supply and sale error data in each sampling interval; The preprocessing sample set is sorted according to the size of the prediction optimization coefficient YH to obtain the prediction optimization sequence of the preprocessing sample set; effectively improving data processing efficiency;

2. In the present invention, after the prediction terminal receives the preprocessed sample set, it splits the preprocessed sample set into three mutually exclusive data sets; uses sequence difference to process the time series and establishes an LSTM neural network model; uses Nguyen-Widrow The method initializes the neural network weights and particle swarm optimization algorithm parameters; inputs the training set, verification set, and test set as historical feature values into the LSTM neural network model for model training, and evaluates the model through the loss function to obtain the overall error of the training sample The smallest optimal short- and medium-term electricity sales data prediction model; improves data prediction accuracy.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is a schematic block diagram of a medium- and short-term electricity sales forecasting method based on deep neural networks according to the present invention.

Detailed ways

The technical solution of the present invention will be described clearly and completely below with reference to the embodiments. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

As shown in Figure 1, a short- and medium-term electricity sales forecast method based on deep neural networks includes the following steps:

Step 1: Obtain the original sample set of each station area; perform data cleaning on the original sample set; preprocess the cleaned data to obtain the preprocessed sample set and cache it to the management center; the original sample set includes date data, climate text data and Historical data of electricity sales;

Among them, data cleaning is performed on the original sample set, including:

Since there may be a small amount of missing data in the historical electricity sales data of the sample station area, and the missing data will affect the VMD decomposition, the present invention uses the average interpolation method to complete the missing data;

Preprocess the cleaned data, including:

Perform unit conversion processing on the historical electricity sales data, complete the sampling time points to ensure their continuity, and use the average interpolation method to fill in the missing data at the sampling points to obtain the Taiwan area electricity sales time series; among them, if the sampling time points and data are missing in large areas, Data from other years during the same period is used for filling;

Step 2: The management center analyzes the prediction optimization coefficient YH of the preprocessing sample sets in each station area, and sorts the preprocessing sample sets according to the size of the prediction optimization coefficient YH to obtain the prediction optimization sequence of the preprocessing sample set; the specific analysis steps are: :

Obtain the station area corresponding to the preprocessed sample set and obtain the power supply area of the station area; count the length of the power supply line in the power supply area as L1, the number of power supply households as HL, and the average power consumption per household as DL;

The regional correlation coefficient GD of the station area is calculated using the formula GD=L1×g2+HL×g3+DL×g4, where g2, g3, and g4 are preset coefficient factors;

Within the preset time period, the supply and sales error data of each sampling interval in the Taiwan area is collected; the supply and sales error data is the difference between power supply and electricity sales; the difference data is taken as a positive number;

Mark the supply and sale error data of each sampling interval as Wi; compare the supply and sale error data Wi with the preset error threshold; among them, the larger the supply and sale error data Wi, the greater the power loss in the corresponding station area, and the higher the Reduced power generation costs; caused additional losses to power grid operation;

The proportion of statistical supply and sale error data Wi > the preset error threshold is Zb. When Wi > the preset error threshold, the difference between Wi and the preset error threshold is obtained and summed to obtain the total excess error value LZ;

Use the formula GS=Zb×g1+LZ×g5 to calculate the power supply loss coefficient GS of the station area, where g1 and g5 are both preset coefficient factors;

Normalize the regional correlation coefficient and power supply loss coefficient and take their values. Use the formula YH=ƒ×(GD×b1+GS×b2) to calculate the prediction optimization coefficient YH of the preprocessed sample set, where b1 and b2 are both is the preset coefficient factor; ƒ is the preset equalization coefficient;

The management center sequentially sends the pre-processed sample sets of each station area to the prediction terminal according to the prediction optimization sequence for electricity sales data prediction model training;

Step 3: After the prediction terminal receives the preprocessed sample set, it splits the preprocessed sample set into three mutually exclusive data sets; the data sets are training set, verification set, and test set; the time series is processed using sequence difference. Establish an LSTM neural network model, in which the number of input nodes of the long short-term memory neural network LSTM is specified based on the number of input variables, the appropriate number of hidden layer nodes, and the number of output nodes representing the predicted value of electricity sales are set;

Among them, sequence difference is used to process time series. The specific steps include:

Based on the monthly and quarterly data of the time series, the 12-month or 4-quarter moving average of electricity sales in Taiwan is obtained, and the long-term moving average trend data Q is obtained;

According to the multiplicative model, calculate Y/Q=S×C×I; where Y represents the year, S represents the seasonal component, C represents the periodic component, and I represents the irregular component;

Recalculate Y/Q based on the same month or quarter of each year and take the average to obtain the arithmetic mean Pi of the same month or quarter of each year;

Using the arithmetic mean Pi of the same month or quarter in each year as the numerator and the sum of the arithmetic mean of all months or quarters as the denominator, the seasonal ratio Si of each month or quarter is calculated, where Si=(N×Pi )/(P1+P2+…+PN); seasonal ratio Si is the correction coefficient of seasonal factors to the long-term trend of electricity sales; P1+P2+…+PN is the sum of the arithmetic average of all months or quarters; N is the number of samples ;

The product of Tt with seasonal factors removed and the corresponding seasonal Si in period t is calculated, which is the predicted value of electricity sales in period t; where Qt is the moving average trend data in period t;

Step 4: Initialize the neural network weights and particle swarm optimization algorithm parameters; conduct model training on the neural network model to obtain the trained short- and medium-term electricity sales data prediction model;

Among them, the Nguyen-Widrow method is used to initialize the neural network weights; the neural network weights include at least one of the input value of the model, the hidden value, the number of model layers, the neuron drop rate, the activation function, the initial value of the feature weight, and the bias. ;

Among them, model training is performed on the neural network model to obtain a trained short- and medium-term electricity sales data prediction model, which specifically includes:

Input the training set, verification set, and test set as historical feature values into the LSTM neural network model for model training, and conduct model evaluation through the loss function to obtain the optimal short- and medium-term electricity sales data prediction model that minimizes the overall error of the training sample;

A further technical solution lies in: model evaluation through loss functions, including:

The root mean square error equation is used as the loss function, and the R2_score coefficient of determination is used as the model fitting evaluation method to evaluate the loss between the model prediction value and the true value;

Determine the trained short- and medium-term electricity sales data prediction model based on the loss assessment results;

Step 5: After obtaining the short- and medium-term electricity sales data prediction model with optimal parameters, use the electricity sales data in the time zone before the time zone to be predicted and the climate text data in the time zone to be predicted as input to the model to obtain the predicted value of electricity sales in the time zone to be predicted.

The above formulas are all numerical calculations after removing the dimensions. The formula is a formula closest to the real situation obtained by collecting a large amount of data for software simulation. The preset parameters and preset thresholds in the formula are set by those skilled in the field according to the actual situation. Or obtain a large amount of data through simulation.

Working principle of the invention:

A short- and medium-term power sales prediction method based on deep neural networks. When working, the original sample set of each station is obtained; the original sample set is data cleaned; the cleaned data is preprocessed to obtain the preprocessed sample set and cached to the management center; the management center conducts prediction optimization coefficient YH analysis on the preprocessed sample sets of each station area, combined with the length of the power supply lines in the power supply area, the number of power supply households, average household electricity consumption, and the supply and sales error data of each sampling interval, Calculate the prediction optimization coefficient YH of the preprocessing sample set; sort the preprocessing sample set according to the size of the prediction optimization coefficient YH to obtain the prediction optimization sequence of the preprocessing sample set; the management center sequentially sorts the preprocessing results of each station area according to the prediction optimization sequence The sample set is sent to the prediction terminal for electricity sales data prediction model training; improving data processing efficiency;

After the prediction terminal receives the preprocessed sample set, it splits the preprocessed sample set into three mutually exclusive data sets; uses sequence difference to process the time series and establishes an LSTM neural network model; uses the Nguyen-Widrow method to initialize the neural network weights values and particle swarm optimization algorithm parameters; input the training set, verification set, and test set as historical feature values into the LSTM neural network model for model training, and conduct model evaluation through the loss function to obtain the optimal medium that minimizes the overall error of the training sample. Short-term electricity sales data prediction model; improve data prediction accuracy.

In the description of this specification, reference to the terms "one embodiment," "example," "specific example," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one aspect of the invention. in an embodiment or example. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The preferred embodiments of the invention disclosed above are only intended to help illustrate the invention. The preferred embodiments do not describe all details, nor do they limit the invention to specific implementations. Obviously, many modifications and variations are possible in light of the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, so that those skilled in the art can better understand and utilize the present invention. The invention is limited only by the claims and their full scope and equivalents.

Claims (7)

1. The medium-short-term electricity sales quantity prediction method based on the deep neural network is characterized by comprising the following steps of:

step one: acquiring an original sample set of each area; performing data cleaning on the original sample set; preprocessing the cleaned data to obtain a preprocessed sample set and caching the preprocessed sample set to a management center; the original sample set comprises date data, climate text data and sales volume history data;

step two: the management center carries out prediction optimization coefficient YH analysis on the pretreatment sample set of each platform region, and the specific analysis steps are as follows:

obtaining a platform region corresponding to a pretreatment sample set, and obtaining a power supply region of the platform region; counting the length of a power supply line in a power supply area to be L1, the number of power supplies to be HL and the average power consumption of users to be DL;

calculating a region association coefficient GD of the station region by using a formula GD=L1×g2+HL×g3+DL×g4, wherein g2, g3 and g4 are preset coefficient factors;

collecting vending error data of each sampling interval of the platform region in a preset time period; the supply and sales error data are difference data of the supply power quantity and the sales power quantity; the difference value data is positive;

marking the vending error data of each sampling interval as Wi; comparing the vending error data Wi with a preset error threshold value; if Wi is larger than a preset error threshold, the power consumption of the corresponding area is large, the power generation cost is increased, and extra loss is caused to the operation of the power grid;

counting the times of the sale error data Wi > the preset error threshold value as Zb, and when Wi > the preset error threshold value, obtaining the difference value of Wi and the preset error threshold value and summing to obtain a super error total value LZ; calculating a power supply loss coefficient GS of the station area by using a formula GS=Zb×g1+LZ×g5, wherein g1 and g5 are preset coefficient factors;

normalizing the area association coefficient and the power supply loss coefficient and taking the values of the area association coefficient and the power supply loss coefficient, and calculating by using a formula YH= ƒ × (GDXb1+GS Xb 2) to obtain a prediction optimization coefficient YH of the pretreatment sample set, wherein b1 and b2 are preset coefficient factors; ƒ is a preset equalization coefficient;

sequencing the pretreatment sample set according to the magnitude of the prediction optimization coefficient YH to obtain a prediction optimization sequence of the pretreatment sample set; then, the pretreatment sample sets of each platform area are sequentially sent to a prediction terminal for electricity selling data prediction model training according to the prediction optimization sequence;

step three: after receiving the pretreatment sample set, the prediction terminal splits the pretreatment sample set into a mutually exclusive training set, a mutually exclusive verification set and a mutually exclusive testing set; processing the time sequence in a sequence difference mode, and establishing an LSTM neural network model; the method comprises the steps of designating the number of LSTM input nodes of a long-short-period memory neural network according to the number of input variables, and setting the number of suitable hidden layer nodes and the number of output nodes representing the predicted value of the sales quantity;

step four: initializing the weight of a neural network and the parameters of a particle swarm optimization algorithm; initializing a neural network weight by adopting an Nguyen-widry method; performing model training on the neural network model to obtain a trained medium-short-term electricity selling data prediction model;

step five: after the middle-short-term electricity selling data prediction model of the optimal parameters is obtained, the electricity selling quantity data of the time zone before the time zone to be predicted and the climate text data of the time zone to be predicted are input as models, and the electricity selling quantity prediction value of the time zone to be predicted is obtained.

2. The method for predicting the medium-short term electricity sales amount based on the deep neural network according to claim 1, wherein the data cleaning of the original sample set comprises the following steps: filling or discarding the null value; performing de-duplication processing on the repeated data; cleaning climate text data with incorrect ranges; and carrying out data verification on the climate text data.

3. The method for predicting the medium-short term electricity sales amount based on the deep neural network according to claim 2, wherein preprocessing the cleaned data comprises the following steps:

carrying out unit conversion processing on the electricity selling history data, complementing sampling time points to ensure continuity of the electricity selling history data, and filling missing data of the sampling points by using an average interpolation method to obtain a time sequence of electricity selling quantity of a platform area; if the sampling time point and the data are missing in a large area, filling is carried out by using the data of other years in the same period.

4. The method for predicting the medium-short-term electricity sales amount based on the deep neural network according to claim 1, wherein the method for processing the time sequence by adopting a sequence difference mode comprises the following specific steps:

according to the month and quarter data of the time sequence, obtaining a moving average value of the sales power of the platform area for 12 months or 4 quarters, and obtaining long-term moving average value trend data Q;

calculating Y/q=s×c×i according to the multiplication model; wherein Y represents year, S represents seasonal component, C represents periodic component, and I represents irregular component;

calculating Y/Q according to the same month or the same quarter of each year again and taking an average value to obtain an arithmetic average Pi of the same month or the same quarter of each year;

calculating to obtain a seasonal ratio Si of each month or quarter by taking an arithmetic mean Pi of the same month or quarter of each year as a molecule and taking the sum of the arithmetic mean of all months or quarters as a denominator, wherein Si= (N×Pi)/(P1+P2+ … +PN); the seasonal ratio Si is a correction coefficient of the seasonal factor to the long-term trend of the sales power; p1+p2+ … +pn is the sum of the arithmetic averages of all months or quarters; n is the number of samples;

calculating to obtain the product of Qt with seasonal factors removed and seasonal Si corresponding to the t period, namely the predicted value of the sales power quantity corresponding to the t period; wherein Qt is moving average trend data for period t.

5. The method for predicting the middle-short term electricity sales based on the deep neural network according to claim 1, wherein the neural network weight comprises at least one of an input value, a hidden value, a model layer number, a neuron discarding rate, an activation function, a characteristic weight initial value and a bias of a model.

6. The method for predicting the medium-short-term electricity sales amount based on the deep neural network according to claim 1, wherein the neural network model is subjected to model training to obtain a trained medium-short-term electricity sales data prediction model, and the method specifically comprises the following steps:

and inputting the training set, the verification set and the test set as historical characteristic values into the LSTM neural network model to perform model training, and performing model evaluation through a loss function to obtain an optimal medium-short term electricity selling data prediction model which minimizes the overall error of the training sample.

7. The method for predicting medium-short term sales power based on deep neural network according to claim 6, wherein the model evaluation by the loss function comprises:

adopting a root mean square error equation as a loss function, adopting an R2_score determining coefficient as a model fitting degree evaluating method, and evaluating the loss condition between a model predicted value and a true value;

and determining a trained medium-short-term electricity selling data prediction model according to the loss evaluation result.