CN104636822B - A kind of resident load prediction technique based on elman neural networks - Google Patents

Abstract

The invention discloses a resident load forecasting method based on elman neural network, which includes: obtaining historical data of resident load and corresponding historical weather parameter data in the previous year; calculating the seasonal index of resident load in each month; Correct the historical load data; determine the input and output data of the neural network, and determine the optimal number of neurons in the hidden layer, so as to establish a neural network based on elman; correct the historical load data of residents and the corresponding historical weather parameters The data is normalized, and then the established neural network is trained to control the prediction error within the preset range; the trained neural network is used to predict the load of residents. The invention has the ability to adapt to time-varying characteristics and seasonal fluctuations of residential loads, can directly predict and reflect the dynamic characteristics of residential loads, has high prediction accuracy, and can be widely used in the field of load prediction of electric power systems.

Description

A Residential Load Forecasting Method Based on Elman Neural Network

technical field

The invention relates to the technical field of power system load forecasting, in particular to a residential load forecasting method based on an elman neural network.

Background technique

As an important energy source, electricity plays a pivotal role in daily life and work. With the rapid development of the national economy, the electricity consumption of the whole society and various industries has also increased steadily. Therefore, the trend of electricity consumption It not only affects the production and operation decision-making and economic benefits of power grid operating enterprises, but also affects the trend analysis of social economy. Reasonable power load forecasting is the precondition for power system to dispatch and plan power resources.

Power loads can generally be divided into industrial loads, commercial loads, residential loads, etc. Among them, industrial loads and commercial loads account for a relatively high proportion of power loads. Power grid companies have always paid more attention to this load forecasting, and have successively built load control systems. and electricity consumption information collection system to complete the data collection and load forecasting of industrial and commercial loads; due to the characteristics of scattered distribution and small scale of residential user loads, the method of centralized forecasting has been adopted all the time, that is, the unit of station area or feeder load Forecasting, the disadvantage of this forecasting method is that the accuracy is not high, especially with the increase of household appliances year by year, the popularity of electric bicycles and the gradual promotion of electric vehicles, the electricity load of residential users shows a steady growth trend and obvious seasonality Fluctuations, the disadvantages of residential user load forecasting through centralized forecasting methods are becoming more and more apparent.

Contents of the invention

In order to solve the above-mentioned technical problems, the purpose of this invention is to provide a kind of resident load forecasting method based on elman neural network.

The technical solution adopted by the present invention to solve its technical problems is:

A resident load forecasting method based on elman neural network, including:

S1. Obtain the historical data of residents' load and the corresponding historical weather parameter data in the previous year, and at the same time divide the valid days in the year into date types;

S2. According to the acquired historical data of resident load, calculate the average number of residents' load in each month, and then calculate the total average value of all averages in the same period, and divide each average number by the total average value to obtain the seasonal index;

S3. Use the seasonal index to correct the historical data of residents' load, divide the historical data of residents' load in each month by the corresponding seasonal index, and obtain the corrected historical data of residents' load;

S4. Determine the input and output data of the neural network, and determine the optimal number of neurons in the hidden layer, thereby establishing a neural network based on elman;

S5. Normalize the corrected historical resident load data and the corresponding historical weather parameter data, and then train the established neural network according to the normalized processed data, and control the prediction error of the neural network to a preset value within the scope;

S6. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then multiply the obtained forecast data by Residential load forecast data is obtained after the seasonal index;

The date type is divided into two types: rest day and working day.

Further, the preset range in the step S5 is 5%-10%.

Further, the historical resident load data includes resident load data for each hour, and the historical weather parameter data includes air temperature, sunshine time and weather type.

Further, the step S4 includes:

S41. Statistically obtain the historical data of resident load, historical weather parameter data and date type, use the resident load data of any day as the output data of the neural network, and at the same time the resident load data of each hour in the week before the day and the The daily weather parameter data and date type are used as the input data of the neural network;

S42. Initialize the neural network, and determine the input node unit vector, the hidden layer node unit vector, the feedback state vector and the output node vector according to the input and output sequences, thereby establishing an elman-based neural network.

Further, the nonlinear state space expression of the neural network based on elman is:

Among them, y(k) represents the m-dimensional output node vector, l(k) represents the m-dimensional hidden layer node unit vector, x(k) represents the u-dimensional input vector, c(k) represents the n-dimensional feedback state vector, w ³ Represents the connection weight from the hidden layer to the output layer, w ² represents the connection weight from the input layer to the hidden layer, w ¹ represents the connection weight from the receiving layer to the hidden layer, g(*) represents the transmission of the output neuron Function, f(*) represents the transfer function of hidden layer neurons.

Further, the step S5 includes:

S51. Perform normalization processing on the corrected historical resident load data and corresponding historical weather parameter data according to the following formula:

Among them, x _k represents the kth parameter value in the historical data sequence of residents' load or the historical weather parameter data sequence, k is a natural number, x _max represents the maximum value in the data sequence where x _k is located, and x _min represents the data sequence where x _k is located the minimum value;

S52. Perform error calculation, weight update, and threshold update on the established neural network according to the normalized data, and then control the prediction error of the elman-based neural network within a preset range.

Further, the elman-based neural network uses the BP algorithm to correct and update the weights, and uses the error sum of squares function to learn the indicator function, and the formula for the indicator function learning is:

In the above formula, E(x) represents the index function, represents the target input vector.

Further, the step S6 includes:

S61. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then obtain the hourly load of the forecast day. forecast data;

S62. Obtain hourly resident load forecast data after multiplying the obtained forecast data by the seasonal index.

Further, after the step S62, the following steps are also included:

S63. After obtaining the actual load data on the forecast day, calculate an error value between the obtained resident load forecast data and the actual load data, and feed back the error value to the neural network.

The beneficial effect of the present invention is: a kind of resident load forecasting method based on elman neural network of the present invention, comprises: obtain last year's resident load historical data and corresponding historical weather parameter data, carry out the valid days number in this year at the same time Date type division; according to the obtained historical data of resident load, calculate the average number of residents' load in each month, and then calculate the total average value of all averages in the same period, and divide each average number in the same period by the total average value to obtain the seasonal index ; Use the seasonal index to correct the historical data of residents' load, divide the historical data of residents' load in each month by the corresponding seasonal index, and obtain the corrected historical data of residents' load; determine the input and output data of the neural network, and determine the optimal The number of neurons in the hidden layer of the hidden layer, so as to establish a neural network based on elman; normalize the corrected historical data of residents' load and the corresponding historical weather parameter data, and then use the normalized data to establish The neural network is trained to control the prediction error of the neural network within the preset range; the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day are taken as the input of the neural network, and the trained neural network is used to The residents' load on the forecast day is predicted, and then the obtained forecast data is multiplied by the seasonal index to obtain the residents' load forecast data. Through the construction of Elman neural network based on seasonal index, this method can predict the residents' load data on the forecast day by combining the historical data of residents' load in relevant areas and the corresponding historical weather parameter data, and can approach any nonlinear mapping with arbitrary precision , does not consider the influence of external noise, has high accuracy, and has the ability to adapt to time-varying characteristics and seasonal fluctuations of resident loads, can directly predict and reflect the dynamic characteristics of resident loads, and has high prediction accuracy. Moreover, this method adds seasonal index features, which can overcome the problems of large seasonal fluctuations of residents' load and incomplete utilization of historical data, and can effectively improve the accuracy and stability of forecast data.

Description of drawings

The present invention will be further described below in conjunction with drawings and embodiments.

Fig. 1 is a kind of flow diagram of the resident's load forecasting method based on elman neural network of the present invention;

Fig. 2 is a structural schematic diagram of a neural network established by a resident load forecasting method based on an elman neural network in the present invention.

Detailed ways

For the convenience of the following description, the following noun explanations are first given:

Elman network: J.L.Elman first proposed it for speech processing in 1990. It is a typical local regression network (global feed for ward local recurrent).

BP algorithm: Error Back Propagation Algorithm, error back propagation algorithm, referred to as BP algorithm.

With reference to Fig. 1, the present invention provides a kind of resident load prediction method based on elman neural network, comprising:

S1. Obtain the historical data of residents' load and the corresponding historical weather parameter data in the previous year, and at the same time divide the valid days in the year into date types;

S2. According to the acquired historical data of resident load, calculate the average number of residents' load in each month, and then calculate the total average value of all averages in the same period, and divide each average number by the total average value to obtain the seasonal index;

S3. Use the seasonal index to correct the historical data of residents' load, divide the historical data of residents' load in each month by the corresponding seasonal index, and obtain the corrected historical data of residents' load;

S4. Determine the input and output data of the neural network, and determine the optimal number of neurons in the hidden layer, thereby establishing a neural network based on elman;

S5. Normalize the corrected historical resident load data and the corresponding historical weather parameter data, and then train the established neural network according to the normalized processed data, and control the prediction error of the neural network to a preset value within the scope;

S6. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then multiply the obtained forecast data by Residential load forecast data is obtained after the seasonal index;

The date type is divided into two types: rest day and working day.

As a further preferred embodiment, the preset range in step S5 is 5%-10%.

As a further preferred embodiment, the historical data of residents' load includes the data of residents' load every hour, and the historical weather parameter data includes air temperature, sunshine time and weather type.

Further as a preferred embodiment, the step S4 includes:

S41. Statistically obtain the historical data of resident load, historical weather parameter data and date type, use the resident load data of any day as the output data of the neural network, and at the same time the resident load data of each hour in the week before the day and the The daily weather parameter data and date type are used as the input data of the neural network;

S42. Initialize the neural network, and determine the input node unit vector, the hidden layer node unit vector, the feedback state vector and the output node vector according to the input and output sequences, thereby establishing an elman-based neural network.

Further as a preferred embodiment, the nonlinear state space expression of the neural network based on elman is:

Among them, y(k) represents the m-dimensional output node vector, l(k) represents the m-dimensional hidden layer node unit vector, x(k) represents the u-dimensional input vector, c(k) represents the n-dimensional feedback state vector, w ³ Represents the connection weight from the hidden layer to the output layer, w ² represents the connection weight from the input layer to the hidden layer, w ¹ represents the connection weight from the receiving layer to the hidden layer, g(*) represents the transmission of the output neuron Function, f(*) represents the transfer function of hidden layer neurons.

Further as a preferred embodiment, the step S5 includes:

S51. Perform normalization processing on the corrected historical resident load data and corresponding historical weather parameter data according to the following formula:

Among them, x _k represents the kth parameter value in the historical data sequence of residents' load or the historical weather parameter data sequence, k is a natural number, x _max represents the maximum value in the data sequence where x _k is located, and x _min represents the data sequence where x _k is located the minimum value;

S52. Perform error calculation, weight update, and threshold update on the established neural network according to the normalized data, and then control the prediction error of the elman-based neural network within a preset range.

Further as a preferred embodiment, the elman-based neural network uses the BP algorithm to correct and update the weights, and uses the error sum of squares function to learn the indicator function, and the formula for learning the indicator function is:

In the above formula, E(x) represents the index function, represents the target input vector.

Further as a preferred implementation manner, the step S6 includes:

S61. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then obtain the hourly load of the forecast day. forecast data;

S62. Obtain hourly resident load forecast data after multiplying the obtained forecast data by the seasonal index.

Further as a preferred implementation manner, after the step S62, the following steps are also included:

S63. After obtaining the actual load data on the forecast day, calculate an error value between the obtained resident load forecast data and the actual load data, and feed back the error value to the neural network.

The present invention will be further described below in conjunction with specific embodiments.

Referring to Figure 1, a resident load forecasting method based on the elman neural network includes:

S1. Obtain the historical data of residents' load and the corresponding historical weather parameter data in the previous year, and at the same time divide the valid days in the year into date types. In this embodiment, the date types are divided into two types: rest days and working days;

S2. According to the acquired historical data of resident load, calculate the average number of residents' load in each month, and then calculate the total average value of all averages in the same period, and divide each average number by the total average value to obtain the seasonal index;

S3. Use the seasonal index to correct the historical data of residents' load, divide the historical data of residents' load in each month by the corresponding seasonal index, and obtain the corrected historical data of residents' load;

S4. Determine the input and output data of the neural network, and determine the optimal number of neurons in the hidden layer, thereby establishing a neural network based on elman; a neural network based on elman includes an input layer, a hidden layer, a succession layer and an output layer , the receiving layer is used to remember the output value of the hidden layer at the previous moment and return the output value to the input of the hidden layer, which increases feedback, is more sensitive to historical data, and is more stable;

S5. Normalize the corrected historical resident load data and the corresponding historical weather parameter data, and then train the established neural network according to the normalized processed data, and control the prediction error of the neural network to a preset value Within the scope; in the present embodiment, preset range is 5%～10%; Carry out error control, can approach arbitrary nonlinear mapping with arbitrary precision;

S6. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then multiply the obtained forecast data by Residential load forecast data is obtained after the seasonal index;

Residential load historical data includes hourly residential load data, and historical weather parameter data includes temperature, sunshine time and weather type.

Specifically, step S4 includes steps S41-S42:

S41. Statistically obtain the historical data of resident load, historical weather parameter data and date type, use the resident load data of any day as the output data of the neural network, and at the same time the resident load data of each hour in the week before the day and the The daily weather parameter data and date type are used as the input data of the neural network; there are 168 load points in the input data, and 24 load points in the output data;

S42. Initialize the neural network, and determine u-dimensional input node unit vector x, n-dimensional hidden layer node unit vector l, n-dimensional feedback state vector c and m-dimensional output node according to the input and output sequence (X, Y) Vector y, thereby establishing a neural network based on elman, the neural network training model established in this embodiment is shown in Figure 2. Among them, X1, X2···Xu are the nodes of the input layer, corresponding to the input weather parameter data of the forecast day, the load and date type of the residents after correction in the previous week; Y1 is the node of the output layer, corresponding to the output forecast day system residents User load; l1, l2...lN are hidden layer nodes, where the number of hidden layer nodes n (that is, the optimal number of hidden layer neurons) is gradually increased by the trial and error method, that is, according to the gradually increased trial C1, C2... CN is the node of the receiving layer, which is used to memorize the output value of the hidden layer unit at the previous moment and return it to the input of the hidden layer.

In this embodiment, the nonlinear state space expression of the neural network based on elman is:

Among them, y(k) represents the m-dimensional output node vector, l(k) represents the m-dimensional hidden layer node unit vector, x(k) represents the u-dimensional input vector, c(k) represents the n-dimensional feedback state vector, w ³ Represents the connection weight from the hidden layer to the output layer, w ² represents the connection weight from the input layer to the hidden layer, w ¹ represents the connection weight from the receiving layer to the hidden layer, g(*) represents the transmission of the output neuron Function, f(*) represents the transfer function of hidden layer neurons, and f(*) generally adopts S function.

Specifically, step S5 includes steps S51-S52:

S51. Using the maximum and minimum method, normalize the corrected historical resident load data and corresponding historical weather parameter data according to the following formula:

Among them, x _k represents the kth parameter value in the historical data sequence of residents' load or the historical weather parameter data sequence, k is a natural number, x _max represents the maximum value in the data sequence where x _k is located, and x _min represents the data sequence where x _k is located the minimum value;

S52. Perform error calculation, weight update, and threshold update on the established neural network according to the normalized data, and then control the prediction error of the elman-based neural network within a preset range.

Due to the particularity of the date type, the rest day is marked as 1 and the working day is marked as 0, which also conforms to the principle of maximum and minimum normalization, that is, normalizing the date type will not affect its specific value.

In step S52, the elman-based neural network uses the BP algorithm to correct and update the weights, and uses the error sum of squares function to learn the indicator function. The formula for learning the indicator function is:

In the above formula, E(x) represents the index function, represents the target input vector.

Specifically, step S6 includes steps S61-S63:

S61. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then obtain the hourly load of the forecast day. Forecast data, that is, obtain the predicted 24 load points;

S62. After multiplying the obtained forecast data by the seasonal index, obtain the hourly resident load forecast data;

S63. After obtaining the actual load data on the forecast day, calculate an error value between the obtained resident load forecast data and the actual load data, and feed back the error value to the neural network. After obtaining the actual load data on the forecast day, compare it with the 24 load points in the forecast data, calculate the error values between the 24 load points and the actual load data, and feed them back to the neural network, so that the neural network training model can be established Make adjustments to make it closer to reality.

The resident load prediction method of the present invention can approach any nonlinear mapping with any precision by establishing a neural network prediction model of the resident load, without considering the influence of external noise, has high precision, and has the characteristics of adapting to time-varying It can directly predict and reflect the dynamic characteristics of residential load, and the prediction accuracy is high. Moreover, this method adds seasonal index features, which can overcome the problems of large seasonal fluctuations of residents' load and incomplete utilization of historical data, and can effectively improve the accuracy and stability of forecast data.

In addition, based on the historical data of residents' electricity consumption and seasonal climate changes, predicting their recent electricity consumption will help residents consciously save electricity and alleviate the urgency of the lack of power resources. At the same time, by combining residents' load forecasting with the use of new energy, it is possible to make full use of various forms of new energy such as solar energy and wind energy, avoid energy waste, and create greater economic benefits for users.

The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments, and those skilled in the art can also make various equivalent deformations or replacements without violating the spirit of the present invention. Equivalent modifications or replacements are all included within the scope defined by the claims of the present application.

Claims (6)

1. A resident load forecasting method based on elman neural network, is characterized in that, comprises: S1. Obtain the historical data of residents' load and the corresponding historical weather parameter data in the previous year, and at the same time divide the valid days in the year into date types; S2. According to the acquired historical data of resident load, calculate the average number of residents' load in each month, and then calculate the total average value of all averages in the same period, and divide each average number by the total average value to obtain the seasonal index; S3. Use the seasonal index to correct the historical data of residents' load, divide the historical data of residents' load in each month by the corresponding seasonal index, and obtain the corrected historical data of residents' load; S4. Determine the input and output data of the neural network, and determine the optimal number of neurons in the hidden layer, thereby establishing a neural network based on elman; S5. Normalize the corrected historical resident load data and the corresponding historical weather parameter data, and then train the established neural network according to the normalized processed data, and control the prediction error of the neural network to a preset value within the scope; S6. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then multiply the obtained forecast data by Residential load forecast data is obtained after the seasonal index; The date type is divided into two types: rest day and working day; The historical data of residents' load includes the residents' load data of each hour, and the historical weather parameter data includes air temperature, sunshine time and weather type; The step S4 includes: S41. Statistically obtain the historical data of resident load, historical weather parameter data and date type, use the resident load data of any day as the output data of the neural network, and at the same time the resident load data of each hour in the week before the day and the The daily weather parameter data and date type are used as the input data of the neural network; S42, the neural network is initialized, and the input node unit vector, the hidden layer node unit vector, the feedback state vector and the output node vector are determined according to the input and output sequence, thereby establishing a neural network based on elman; The neural network based on elman adopts BP algorithm to carry out weight correction and update, and adopts error sum of squares function to carry out index function learning, and the formula of described index function learning is: In the above formula, E(x) represents the index function, represents the target input vector. 2. A kind of resident load forecasting method based on elman neural network according to claim 1, characterized in that, the preset range in the step S5 is 5% to 10%. 3. a kind of resident load forecasting method based on elman neural network according to claim 1, is characterized in that, the nonlinear state space expression of the neural network based on elman is: Among them, y(k) represents the m-dimensional output node vector, l(k) represents the m-dimensional hidden layer node unit vector, x(k) represents the u-dimensional input vector, c(k) represents the n-dimensional feedback state vector, w ³ Represents the connection weight from the hidden layer to the output layer, w ² represents the connection weight from the input layer to the hidden layer, w ¹ represents the connection weight from the receiving layer to the hidden layer, g(*) represents the transmission of the output neuron Function, f(*) represents the transfer function of hidden layer neurons. 4. a kind of resident load forecasting method based on elman neural network according to claim 1, is characterized in that, described step S5, comprises: S51. Perform normalization processing on the corrected historical resident load data and corresponding historical weather parameter data according to the following formula: Among them, x _k represents the kth parameter value in the historical data sequence of residents' load or the historical weather parameter data sequence, k is a natural number, x _max represents the maximum value in the data sequence where x _k is located, and x _min represents the data sequence where x _k is located the minimum value; S52. Perform error calculation, weight update, and threshold update on the established neural network according to the normalized data, and then control the prediction error of the elman-based neural network within a preset range. 5. a kind of resident load forecasting method based on elman neural network according to claim 4, is characterized in that, described step S6, comprises: S61. Obtain the historical data of residents' load one week before the forecast day, the weather parameter data and date type of the forecast day as the input of the neural network, use the trained neural network to predict the residents' load on the forecast day, and then obtain the hourly load of the forecast day. forecast data; S62. Obtain hourly resident load forecast data after multiplying the obtained forecast data by the seasonal index. 6. a kind of resident load forecasting method based on elman neural network according to claim 5, is characterized in that, also comprises the following steps after described step S62: S63. After obtaining the actual load data on the forecast day, calculate an error value between the obtained resident load forecast data and the actual load data, and feed back the error value to the neural network.