CN110570023B - Short-term commercial power load prediction method based on SARIMA-GRNN-SVM - Google Patents

Abstract

The invention relates to a short-term commercial power load prediction method based on SARIMA-GRNN-SVM, which comprises the following steps: obtaining influence factors of commercial power load fluctuation through analysis of a load curve; constructing a single prediction model for the commercial power load time sequence prediction and the multi-factor regression prediction; constructing an SVM model, carrying out parameter optimization and training on the SVM model by utilizing a training sample, and carrying out parameter optimization on the SVM model by a grid search and cross verification method; and inputting the predicted value of the predicted day obtained by the SARIMA model and the GRNN model into the trained SVM model to obtain the predicted value of the commercial power load of the predicted day. The method solves the problem that the single prediction model cannot comprehensively consider the periodic change of the commercial load and influence factors to cause the prediction result to be easy to generate larger errors, and improves the prediction accuracy and the robustness.

Description

A Short-Term Commercial Power Load Forecasting Method Based on SARIMA-GRNN-SVM

technical field

The invention belongs to the technical field of electric vehicle components, in particular to a short-term commercial power load forecasting method based on SARIMA-GRNN-SVM.

Background technique

Short-term power load forecasting is the basis for power supply enterprises to formulate power generation and dispatch plans and ensure the stable and reliable operation of the power system. In the context of the rapid development of the power spot market, short-term power load forecasting can enable power supply companies to grasp the power consumption situation in a timely manner, formulate reasonable power prices, flexibly adjust power sales strategies, and improve transaction efficiency and economic benefits in the power spot market.

In recent years, due to the adjustment of my country's industrial structure and the improvement of residents' living standards, the commercial power load has a clear upward trend. Compared with industrial loads and residential loads, commercial power loads are more controllable, and the difference between peaks and valleys is more obvious, making them suitable for trading in the power spot market. It can be seen that the short-term forecast of commercial power load is of great significance to the safe and economic operation of the power grid and the smooth development of the power spot market. At this stage, there are few studies on commercial power load forecasting, most of which focus on mid- and long-term commercial power load forecasting, and do not consider the impact of factors such as weather and temperature on commercial power loads, making it difficult to meet the needs of the power spot market for short-term commercial power load forecasting. Require.

Therefore, based on these problems, a short-term commercial power load forecasting method based on SARIMA-GRNN-SVM is provided to overcome the problem that a single forecasting model cannot comprehensively consider the periodic changes and influencing factors of commercial loads, resulting in large errors in forecasting results. has important practical significance.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a SARIMA-GRNN-SVM-based short-term forecasting model that overcomes the problem that a single forecasting model cannot comprehensively consider the periodic changes and influencing factors of commercial loads, resulting in large errors in forecasting results. Commercial Electric Load Forecasting Methods.

The present invention solves its technical problem and realizes by taking the following technical solutions:

A short-term commercial power load forecasting method based on SARIMA-GRNN-SVM, comprising the following steps:

S1. Collect recent local commercial power load data, and draw the average load curve or/and the load curve of a certain period of time during this period, and obtain the influencing factors of commercial power load fluctuations through the analysis of the load curve;

S2. Construct a single forecasting model for commercial power load time series forecasting and multi-factor regression forecasting;

S201. Establish a SARIMA model to realize the prediction of commercial power load time series;

To establish the SARIMA model, the formula is as follows:

in:

θ(B)＝1-θ ₁ B-…-θ _q B ^q

Φ(B ^s )＝1-Φ ₁ B ^S -…-Φ _P B ^S

Θ(B ^s )＝1-Θ ₁ B ^S -…-Θ _Q B ^S

为差分后的平稳时间序列；B表示滞后算子，

表示差分算子；

为季节自回归模型，

为p阶自回归多项式，

为非季节自回归参数；Φ(B^s)表示季节自回归多项式，Φ₁，Φ₂，…Φ_P为P阶季节自回归参数；Θ(B^S)θ(B)表示季节移动平均模型，其中θ(B)表示q阶移动平均多项式，θ₁，θ₂，…θ_q为非季节移动平均参数，Θ(B^S)表示季节移动平均多项式，Θ₁，Θ₂，…Θ_Q为Q阶季节移动平均参数，ε(t)为高斯噪音；In the formula, x(t) is the commercial power load time series,

is the stationary time series after difference; B represents the lag operator,

Represents the difference operator;

is a seasonal autoregressive model,

is an autoregressive polynomial of order p,

is a non-seasonal autoregressive parameter; Φ(B ^s ) represents a seasonal autoregressive polynomial, Φ ₁ , Φ ₂ , ... Φ _P is a P-order seasonal autoregressive parameter; Θ(B ^S )θ(B) represents a seasonal moving average model, Where θ(B) represents the q-order moving average polynomial, θ ₁ , θ ₂ , ... θ _q are non-seasonal moving average parameters, Θ(B ^S ) represents the seasonal moving average polynomial, Θ ₁ , Θ ₂ , ... Θ _Q is Q order seasonal moving average parameter, ε(t) is Gaussian noise;

Establish the SARIMA model, take the historical load and time series of a continuous period of time before the forecast date as input, use the Augmented Dickey-Fuller test to stabilize the original time series and determine the order of difference, and use the autocorrelation coefficient and partial autocorrelation coefficient graph to determine the possible The model parameters of Akaike's information criterion and Bayesian information criterion are used to filter and output the forecast value of the forecast day;

S202. Using the GRNN model to perform multi-factor regression prediction on commercial power loads

The GRNN model is constructed, and the influencing factors of commercial power load fluctuations determined in step S1 of a continuous period of time before the forecast date and the load at the same time of the same date type as the input are used as input, and the training method of circular cross-validation is used to make the predicted value of the output the most excellent;

S3. Construct the SVM model, and use the training samples to optimize and train the parameters of the SVM model, and optimize the parameters of the SVM model through the grid search and cross-validation methods; among them, for the forecast of the working day, select the single forecast of the first three working days The model prediction value and actual load value are used as training samples; for the rest day prediction, the single prediction model prediction value and load actual value of the two days before the prediction day and the previous day are selected as training samples; the single prediction model prediction value is used The predicted value obtained by SARIMA model and GRNN model in steps S201 and S202;

S4. Input the predicted value of the predicted day obtained by the SARIMA model and GRNN model in steps S201 and S202 into the trained SVM model in step S3, and obtain the predicted value of the commercial power load on the predicted day.

Further, in the step S3, before training the SVM model, the Raida criterion is used to judge the forecast date in the training sample. If the forecast result of the forecast date is judged to be an abnormal value, the forecast date is eliminated, and the training sample is sent to The previous day was postponed.

Further, the local commercial power load data collected in the step S1 is at least one year old.

Further, the GRNN model in the step S202 is composed of an input layer, a pattern layer, a summation layer and an output layer; each neuron in the input layer directly transmits the input data to the pattern layer, and the pattern layer transfers the samples through the radial transfer function It is transmitted to the summation layer, and the summation layer sums the two algorithms and transmits the data to the output layer, and the output layer uses a linear function to output the result.

Further, the radial transfer function of the mode layer is:

In the formula, X is the input variable of the network; _Xi is the learning sample corresponding to the i-th neuron; σ represents the smooth factor, and the optimal σ is found by using the training method of circular cross-validation.

Further, the summation layer utilizes the following two summation methods: one is the weighted sum of the output of each neuron in the calculation mode layer; the other is the sum of the outputs of each neuron in the calculation mode layer; the two types of formulas are respectively :

In the formula, j=1,2,...,m; h _ij is the jth element in the dependent variable of the i-th training sample.

Advantage and positive effect of the present invention are:

The SARIMA model adopted in the present invention can describe the sequence showing periodic characteristic time by seasonal variation, and when dealing with the time series problem, the correlation relationship between data and the interference of random factors can be considered at the same time, and the prediction accuracy is high; the GRNN model is An improved algorithm of radial basis network has high calculation accuracy and short time required; the SVM model adopted in the present invention utilizes training sample data to construct an optimal hyperplane in a high-dimensional space, and has strong generalization ability. Fewer samples are required; the present invention overcomes the problem that a single forecasting model cannot comprehensively consider the periodic changes and influencing factors of commercial loads, resulting in large errors in the forecasting results, and improves the forecasting accuracy and robustness.

Description of drawings

The technical solutions of the present invention will be described in further detail below in conjunction with the drawings and embodiments, but it should be known that these drawings are only designed for the purpose of explanation, and therefore are not intended to limit the scope of the present invention. Furthermore, unless otherwise indicated, the drawings are only intended to conceptually illustrate the architectural configurations described herein and are not necessarily drawn to scale.

Fig. 1 is the 2017 average daily load curve of a shopping and leisure center provided by Embodiment 1 of the present invention;

Fig. 2 (a) is the average load curve of a certain shopping and leisure center in January 2017 during the hour of business hours provided by Embodiment 1 of the present invention;

Fig. 2 (b) is the hourly average load curve of a certain shopping and leisure center in April 2017 during the business hours provided by Embodiment 1 of the present invention;

Fig. 2 (c) is the average load curve of a shopping and leisure center in July 2017 during the whole hour of business hours provided by Embodiment 1 of the present invention;

Fig. 2 (d) is the average load curve of a shopping and leisure center in October 2017 during the whole hour of business hours provided by Embodiment 1 of the present invention;

Fig. 3 is the forecasted daily business hours load actual value and single item model forecast value that the embodiment 1 of the present invention provides;

Fig. 4 is the actual load actual value and the SVM model prediction value of forecast day business hours provided by embodiment 1 of the present invention;

Detailed ways

It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

Example 1

Fig. 1 is the 2017 average daily load curve of a certain shopping and leisure center provided by Embodiment 1 of the present invention; Fig. 2 (a) is the hourly average load of a certain shopping and leisure center provided by Embodiment 1 of the present invention in January 2017 on the whole point of business hours Curve; Fig. 2 (b) is the hourly average load curve of a certain shopping and leisure center in April 2017 during the business hours provided by Embodiment 1 of the present invention; Fig. 2 (c) is a certain shopping and leisure center provided by Embodiment 1 of the present invention The hourly average load curve of the business hours in July, 2017; Fig. 2 (d) is the average load curve of a certain shopping and leisure center in October 2017 during the hourly hours of business hours provided by Embodiment 1 of the present invention; Fig. 3 is the average load curve of the hour of the present invention The actual value of the forecasted business hours load and the single item model forecast value provided in Embodiment 1; FIG. 4 is the actual value of the forecasted business hours load and the SVM model forecast value provided by Embodiment 1 of the present invention;

According to Figures 1, 2(a), 2(b), 2(c), and 2(d): through the analysis of the load curve, the influencing factors of commercial power load fluctuations are: weather, maximum temperature, and minimum temperature , date type;

Commercial power load forecasting for a large leisure shopping center from July 5, 2017 to July 10, 2017. The actual value of commercial power load during the business hours of 28 days before each forecast date and the weather, maximum temperature, minimum temperature, date type, and load at the same time of the same date type as input for the 14 days before each forecast date were used as input to establish a SARIMA and GRNN single forecast model , and the predicted results are shown in Figure 3.

It can be seen from Figure 3 that the predicted values of each single forecasting model are roughly consistent with the actual value of the commercial power load, but the accuracy of the single forecasting model's prediction results fluctuates greatly, which shows that its robustness is poor.

Utilize the method described in the present invention to carry out parameter optimization and training to SVM: carry out the parameter optimization of SVM model through grid search and cross-validation method; The actual load value is used as a training sample; for the rest day prediction, the single prediction model prediction value and the actual load value of the two days before the prediction day are selected as the same date type day as the training sample; among them, the single prediction model prediction value adopts SARIMA model, GRNN the predicted value from the model;

After optimizing and training the parameters of the SVM, input the prediction results of the single prediction model from July 5 to July 10, 2017 into the SVM to obtain the prediction value of the combined prediction model, as shown in Figure 4.

It can be seen from Fig. 4 that the combined forecasting model proposed by the present invention fits the changing trend of the commercial power load well, and the forecasted result is close to the actual value, except that the forecasted value differs greatly from the actual value on July 6. After drawing After the IDA criterion was detected, there were 5 outliers, so the data on July 6 will no longer be used in subsequent forecasts. Specifically, taking the errors e ₁ , e ₂ , ..., e ₃₆ at each time of the three forecast days in the training sample as the judgment standard, calculate the average E and standard deviation σ; calculate the error v ₁ at different times of the judged day , v ₂ ,...v ₁₂ and the difference V _n (n=1,2,...12) of the judgment standard average value E; if |V _n |＞3σ, it is considered that the forecast value at this moment is an abnormal value; if the forecast date If there are more than 5 outliers, the data of that day will no longer be used for subsequent forecasts.

In order to measure the accuracy and robustness of the prediction model, the present embodiment adopts mean absolute percentage error (MAPE) and root mean square error (RMSE) as error standards, respectively as shown in the following formula:

In the formula, n is the number of prediction time.

Using the above formula to obtain the average absolute percentage error Em and root mean square error Er of the single model and combined model from July 5 to July 10, 2017, as shown in Table 1:

Table 1 Comparison of errors between combined forecasting model and single forecasting model

It can be seen from Table 1 that the MAPE and RMSE of the combined forecasting model proposed in this example are both smaller than the single forecasting model, indicating that the combined forecasting model combines the advantages of the single forecasting model, extracts more comprehensive information, forecasting accuracy and robustness better than a single predictive model.

Example 2

Commercial power load forecasting is carried out for another commercial complex during the business hours from January 13 to January 19, 2017. The error comparison between the combined model and the single model is shown in Table 2.

Table 2 Comparison of errors between combined forecasting model and single forecasting model

It can be seen from Table 2 that because the heating load in winter is smaller than the cooling load in summer, the fluctuation of commercial power load in winter is smaller than that in summer, and the prediction accuracy of single forecasting model and combined forecasting model are better than summer forecasting. And it can also be obtained that the MAPE and RMSE of the combined forecasting model proposed in this embodiment are smaller than each single forecasting model, and the forecasting accuracy and robustness are better than the single forecasting model, which can meet the short-term forecasting accuracy and cycle requirements of commercial power loads.

The above embodiments have described the present invention in detail, but the content described is only a preferred embodiment of the present invention, and cannot be considered as limiting the implementation scope of the present invention. All equivalent changes and improvements made according to the application scope of the present invention shall still belong to the scope covered by the patent of the present invention.

Claims (5)

1. A short-term commercial power load forecasting method based on SARIMA-GRNN-SVM, characterized in that: comprise the steps: S1. Collect recent local commercial power load data, and draw the average load curve or/and the load curve of a certain period of time during this period, and obtain the influencing factors of commercial power load fluctuations through the analysis of the load curve; S2. Construct a single forecasting model for commercial power load time series forecasting and multi-factor regression forecasting; S201. Establish a seasonal autoregressive differential moving average SARIMA model to realize the prediction of commercial power load time series; To establish the SARIMA model, the formula is as follows:

in:

θ(B)=1-θ ₁ B-…-θ _q B ^q Φ(B ^s )＝1-Φ ₁ B ^S -…-Φ _P B ^S Θ(B ^s )＝1-Θ ₁ B ^S -…-Θ _Q B ^S

In the formula, x(t) is the commercial power load time series,

is the stationary time series after difference; B represents the lag operator,

Represents the difference operator;

is a seasonal autoregressive model,

is an autoregressive polynomial of order p,

is a non-seasonal autoregressive parameter; Φ(B ^s ) represents a seasonal autoregressive polynomial, Φ ₁ , Φ ₂ , ... Φ _P is a P-order seasonal autoregressive parameter; Θ(B ^S )θ(B) represents a seasonal moving average model, Where θ(B) represents the q-order moving average polynomial, θ ₁ , θ ₂ , ... θ _q are non-seasonal moving average parameters, Θ(B ^S ) represents the seasonal moving average polynomial, Θ ₁ , Θ ₂ , ... Θ _Q is Q order seasonal moving average parameter, ε(t) is Gaussian noise; Establish the SARIMA model, take the historical load and time series of a continuous period of time before the forecast date as input, use the Augmented Dickey-Fuller test to stabilize the original time series and determine the order of difference, and use the autocorrelation coefficient and partial autocorrelation coefficient graph to determine the possible The model parameters of Akaike's information criterion and Bayesian information criterion are used to filter and output the forecast value of the forecast day; S202. Using the generalized regression neural network GRNN model to perform multi-factor regression prediction on commercial power loads The GRNN model is constructed, and the influencing factors of commercial power load fluctuations determined in step S1 of a continuous period of time before the forecast date and the load at the same time of the same date type as the input are used as input, and the training method of circular cross-validation is used to make the predicted value of the output the most excellent; The GRNN model is composed of an input layer, a pattern layer, a summation layer and an output layer; each neuron in the input layer directly transmits the input data to the pattern layer, and the pattern layer transmits samples to the summation layer through a radial transfer function, adding The sum layer sums the two algorithms and transmits the data to the output layer, and the output layer uses a linear function to output the result; S3. Construct the SVM model, and use the training samples to optimize and train the parameters of the SVM model, and optimize the parameters of the SVM model through the grid search and cross-validation methods; among them, for the forecast of the working day, select the single forecast of the first three working days The model prediction value and actual load value are used as training samples; for the rest day prediction, the single prediction model prediction value and load actual value of the two days before the prediction day and the previous day are selected as training samples; the single prediction model prediction value is used The predicted value obtained by SARIMA model and GRNN model in steps S201 and S202; S4. Input the predicted value of the predicted day obtained by the SARIMA model and GRNN model in steps S201 and S202 into the trained SVM model in step S3, and obtain the predicted value of the commercial power load on the predicted day. 2. A kind of short-term commercial power load forecasting method based on SARIMA-GRNN-SVM according to claim 1, characterized in that: in the step S3, before training the SVM model, use the Raida criterion for training samples If it is judged that the prediction result of the prediction day is an abnormal value, the prediction day will be eliminated, and the training sample will be postponed by one day. 3. A short-term commercial power load forecasting method based on SARIMA-GRNN-SVM according to claim 1, characterized in that: the local commercial power load data collected in the step S1 is at least one year old. 4. a kind of short-term commercial power load forecasting method based on SARIMA-GRNN-SVM according to claim 1, is characterized in that: the radial transfer function of described model layer is:

In the formula, X is the input variable of the network; _Xi is the learning sample corresponding to the i-th neuron; σ represents the smooth factor, and the optimal σ is found by using the training method of circular cross-validation. 5. A kind of short-term commercial power load forecasting method based on SARIMA-GRNN-SVM according to claim 4, characterized in that: the summation layer utilizes the following two summation methods: one is that each neuron of the calculation model layer The weighted sum of unit output; the other is the sum of the output of each neuron in the calculation mode layer; the two types of formulas are:

In the formula, j=1,2,...,m; h _ij is the jth element in the dependent variable of the i-th training sample.

Images