Medium-term load prediction method based on feedforward neural network and time sequence

Abstract

The invention discloses a medium-term load prediction method based on a feedforward neural network and a time sequence. The medium-term load prediction method comprises the following steps: step 1, establishing an in-out power basic load prediction model; 2, establishing a prediction correction model for the influence of the temperature accumulation effect on the load; step 3, establishing a prediction correction model for the influence of humidity change on the load; 4, establishing a meteorological sensitive load prediction model after considering the change relation of humidity, temperature and load; 5, establishing a total load prediction model; and 6, evaluating the precision of the model by adopting sample conformance data. According to the invention, weather prediction information and holiday and festival information are incorporated into the load prediction model, and the load prediction model is established based on the time sequence method and the feedforward neural network, so that the precision of load prediction is improved.

Description

Technical Field

[0001] The invention relates to a medium-term load forecasting method based on a feedforward neural network and a time series for the field of power load forecasting. Background Art

[0002] Since it is difficult to obtain accurate weather forecasts in the future, how to deal with meteorological information has become one of the most controversial issues in load forecasting. So far, researchers have implemented three approaches: 1) completely excluding weather from the load forecasting model; 2) assuming that future weather information is completely known; 3) integrating weather forecasts into the load forecasting model.

[0003] Regarding the time scale of load forecasting, each time scale has different methods and practical applications. Ultra-short-term forecasts are usually a few hours in advance, which is crucial for power balance and real-time control of the power system; short-term forecasts are generally 1 day or a few days in advance, which is crucial for power system operation and scheduling and short-term price forecasts. The medium-term forecast time scale ranges from 1 day to several months or even a year, which can provide information for system operation and scheduling, medium- and long-term transactions, maintenance, etc. The long-term forecast time scale can range from one year to several decades, and is mainly used to guide power system planning.

[0004] At present, there are many studies on short-term load forecasting, but relatively few on medium-term forecasting. The most controversial issue is how to deal with weather information, especially since it is extremely difficult to make accurate weather forecasts for the next few days or weeks, and the Meteorological Bureau is usually unable to provide such accurate forecasts. In response to this, most studies do not consider weather in medium-term load forecasting models, or incorporate future weather data into the forecasting model to avoid this problem. Since the impact of weather forecast errors on load forecasting is not considered, the forecasting performance cannot meet the requirements. The third strategy is to include weather forecast data in the load forecasting model, rather than actual future weather information, which makes it difficult for the forecasting model to handle the uncertainty of weather information, and also further increases the uncertainty of the forecast results. How to establish a load forecasting model that incorporates future weather forecast data to improve the accuracy of medium-term load forecasting is the main goal of technicians. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a medium-term load forecasting method based on a feedforward neural network and time series, incorporate meteorological forecast information and holiday information into the load forecasting model, and establish a load forecasting model based on a time series method and a feedforward neural network to improve the accuracy of load forecasting.

[0006] A technical solution to achieve the above object is: a medium-term load forecasting method based on a feedforward neural network and a time series, comprising the following steps:

[0007] Step 1: Establish an inbound and outbound power base load prediction model;

[0008] Step 2, establishing a prediction and correction model for the impact of temperature accumulation effect on load;

[0009] Step 3: Establish a prediction and correction model for the impact of humidity changes on load;

[0010] Step 4, establish a meteorological sensitive load prediction model after considering the relationship between humidity, temperature and load changes;

[0011] Step 5, establishing a total load forecasting model;

[0012] Step 6: Use sample matching data to evaluate the accuracy of the model.

[0013] Furthermore, step 1, establishing an in-and-out power base load forecasting model is specifically: using linear regression and multi-layer feedforward neural network methods to establish a medium-term power load forecasting model.

[0014] Furthermore, a multi-layer feedforward neural network is selected to construct a multi-layer artificial neural network model, including an input layer, a hidden layer and an output layer. The specific model is:

[0015]

[0016] Where: y t represents the output variable at time t, X represents the vector including the lagged output variable and the external variable input, n represents the number of input variables, and h represents the number of neurons in the hidden layer; f() and g() are S-type nonlinear activation functions; in addition, v j and w ij are weight parameters, representing the connection strength between hidden nodes and output and between input nodes and hidden nodes, b and b j is the bias parameter for the intercept in the linear model.

[0017] Furthermore, the model prediction accuracy and performance were evaluated by calculating the root mean square error and mean absolute percentage error;

[0018]

[0019] RMSE is the root mean square error, and N is the number of prediction accuracy comparisons.

[0020] Furthermore, in step 2, the prediction and correction model for the impact of temperature accumulation effect on load is specifically as follows:

[0021]

[0022] Where: T is the corrected temperature value; i represents the i-th hour before the predicted time, T irepresents the temperature value one hour before the predicted time; T jd represents the temperature value j days before the forecast date; m i , nj correspond to the temperature sensitive load change and human perception hysteresis correction coefficient respectively.

[0023] Furthermore, in step 3, the prediction and correction model for the impact of humidity changes on load is specifically as follows:

[0024]

[0025] Where: H is the humidity value of the forecast period after correction; H0 is the humidity value of the forecast period before correction; H i is the humidity value of the i-th period before the prediction; k i is the weight parameter; P is taken as 1, which means that only the humidity values ​​in the first two periods of the forecast are taken into account.

[0026] Furthermore, in step 4, after considering the relationship between humidity, temperature and load changes, the meteorological sensitive load prediction model is specifically as follows:

[0027]

[0028] Where: L w is the meteorological sensitive load; C is a constant; H is the relative humidity value obtained after correlation correction; T h is the maximum temperature during the forecast period; T a is the average temperature during the prediction period, and k1, k2, k3, k4, and k5 are weight parameters.

[0029] Furthermore, in step 5, the total load forecasting model is specifically as follows:

[0030] L=L i +L w

[0031] Where: L is the total load value, L i is the basic load value, L w is the corrected meteorological sensitive load value.

[0032] Furthermore, in step 6, the Copula function is used to analyze the nonlinear relationship between variables, and the influencing factor analysis is evaluated on the model correlation analysis to obtain the factors with strong correlation.

[0033] Furthermore, the factors with strong correlation are: maximum temperature, minimum temperature, average temperature, humidity, and holiday type.

[0034] The present invention is a medium-term load forecasting method based on a feedforward neural network and time series. Compared with the existing model, the model that incorporates complete future weather information is significantly better than the model that does not consider weather data, and the prediction results are good within the above time range. The model has a high accuracy of 1-7 days prediction, and the subsequent accuracy gradually decreases. The accuracy of 1 week prediction is close to that of 6 months prediction. For a shorter prediction time range, the autoregressive model of the present invention has obvious advantages; the neural network model prediction of the present invention is better than the linear regression model in both short and long time ranges. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Output prediction curve for FNN model;

[0036] Figure 2 is the FNN prediction error curve;

[0037] Figure 3 This is a comparison chart of holiday forecast results;

[0038] Figure 4 A comparison chart of forecast results for working days;

[0039] Figure 5 It is a comparison chart of load forecast curve;

[0040] Figure 6 This is the prediction result graph of the multiple linear regression model. DETAILED DESCRIPTION

[0041] In order to better understand the technical solution of the present invention, the following is a detailed description through specific embodiments:

[0042] A medium-term load forecasting method based on a feedforward neural network and a time series of the present invention comprises the following steps:

[0043] Step 1: Establish an inbound and outbound power base load prediction model;

[0044] Step 2, establishing a prediction and correction model for the impact of temperature accumulation effect on load;

[0045] Step 3: Establish a prediction and correction model for the impact of humidity changes on load;

[0046] Step 4, establish a meteorological sensitive load prediction model after considering the relationship between humidity, temperature and load changes;

[0047] Step 5, establishing a total load forecasting model;

[0048] Step 6: Use sample matching data to evaluate the accuracy of the model.

[0049] Step 1, select a multi-layer feedforward neural network (FNN) and build a multi-layer artificial neural network model: input layer, hidden layer and output layer. In FNN, information is only passed forward from input nodes, hidden nodes and output nodes. The input layer obtains data and passes it to the hidden layer using the weights between the input layer and the hidden layer. The neurons in the hidden layer process the data and pass it to the output layer using the weights between the hidden layer and the output layer. Finally, the neural network in the output layer processes the data again and outputs the result. The FNN mathematical model is as follows:

[0050]

[0051] Where: y t represents the output variable at time t, X represents the vector including the lagged output variable and the external variable input, n represents the number of input variables, and h represents the number of neurons in the hidden layer; and is a S-shaped nonlinear activation function.

[0052] In addition, v j and w ij are weight parameters, representing the connection strength between hidden nodes and output and between input nodes and hidden nodes, b and b j is a bias parameter similar to the intercept in a linear model.

[0053] When using the FNN model for load forecasting, forecasts can be carried out in different scenarios. Scenario 1: Fit a simple model in FNN and make forecasts based only on load data without external characteristic variables; Scenario 2: Includes monthly (M j ,t), weekly (W i ,t) and holidays (H t ) characteristic variables are integrated into the load data for prediction; Scenario 3: Load and temperature variables (temp t and d 35 (temp t -35)) and characteristic variables such as Wi,t and Ht are integrated to carry out load forecasting; Scenario 4: including external influencing factors, temperature, d 35 (temp t -35), W i,t , H t and M j,t Various characteristic variables such as load variables are integrated into the load data for prediction.

[0054] The model prediction accuracy is compared by calculating the root mean square error (RMSE):

[0055]

[0056] Where N is the number of data observations in the window for prediction accuracy comparison, l t , k is the load predicted at time t. In addition, RMSE and mean absolute percentage error (MAPE) are used to evaluate the prediction performance considering weather information. MAPE is defined as:

[0057]

[0058] Step 2 and step 3, medium-term load forecasting correction model considering meteorological factors.

[0059] Meteorological information is added to load forecasting to improve forecast accuracy, especially in medium-term and short-term load forecasting. Meteorological data such as humidity and temperature are commonly used in the load forecasting process. Some of them are consistent with the load change trend, and some are even opposite, making meteorological information correction more important in the construction of forecasting models. Among many influencing factors, temperature and humidity are the focus of meteorological factor correction.

[0060] Step 2. Usually, the load will have a cumulative effect due to the continuous increase in temperature, which makes its change trend difficult to predict during the temperature change process, making it more important to analyze the load change trend under the temperature cumulative effect. When the temperature changes significantly in the forecast period relative to the previous period, load hysteresis will occur, mainly due to the change of temperature-sensitive loads and the hysteresis of human perception. Therefore, load forecast corrections can be implemented under different temperature change conditions, namely the impact of the cumulative effect of temperature on load changes in the previous i periods and the previous j days. The specific model is as follows:

[0061]

[0062] Where: T is the corrected temperature value; i represents the i-th hour before the predicted time, T i represents the temperature value one hour before the predicted time; T jd represents the temperature value j days before the forecast date; m i 、n j The correction coefficients corresponding to the above two cases are based on the temperature accumulation effect as the basic condition.

[0063] Step 3. Humidity and power load are usually negatively correlated. Considering humidity changes during the forecast period can help improve load forecasting accuracy. The specific correction model is as follows:

[0064]

[0065] Where: H is the humidity value of the forecast period after correction; H0 is the humidity value of the forecast period before correction; H i is the humidity value of the i-th period before the prediction; k iis the parameter to be determined; P is taken as 1, which means that only the humidity values ​​in the first two time periods are considered.

[0066] Step 4. After considering the relationship between humidity, temperature and load changes, when conducting load forecasting, we need to focus on two points: real-time temperature and humidity changes, and meteorological sensitive loads.

[0067]

[0068] Where: L w is the meteorological sensitive load; C, H, T h , T a are the constant, the relative humidity value after correlation correction, the maximum temperature during the forecast period, and the average temperature during the forecast period. k1, k2, k3, k4, and k5 are related parameters to be determined.

[0069] The predicted load curve is corrected with the help of the above formula, and combined with the meteorological sensitive load, it is superimposed on the basic load to obtain the predicted value of the total load:

[0070] L=L i +L w Where: L is the total load value, L i is the basic load value, L w is the corrected meteorological sensitive load value.

[0071] During the load forecasting process, the input and output numbers need to be normalized.

[0072]

[0073] After the load collection data is obtained, it needs to be analyzed and preprocessed. After removing the "bad data", the prediction model can be trained. Among them, the number of samples must be reasonable to avoid the prediction model being incomplete due to too few samples, which affects the accuracy of parameter selection. Too many samples will increase the training time.

[0074] The most important factor affecting medium-term load forecasting is temperature. The following will compare the prediction performance of the models in different time periods: a) short-term: from 1 day; b) medium-term: 10 days to 1 month in advance. The 2022-2024 dataset is divided into two parts: one-year dataset is used for estimation, and the prediction performance of different models in 2022-2023 is compared; a one-year rolling window is used to estimate all models, that is, a window from time t*-364-J to t*-J, where t* represents the date you want to predict and J is the number of steps forward corresponding to the prediction range (for example, J=1 is used for the prediction of the previous day, J=2 is used for the prediction of the previous two days, J=30 is used for the prediction of the previous month, etc.). For example, the model used for a one-day-ahead forecast for January 1, 2024 is estimated based on data from January 1, 2023 to December 31, 2023, and the model used for a two-day-ahead forecast for January 31, 2024 is estimated based on data from December 31, 2022 to December 31, 2023.

[0075] For the six-months-ahead forecast to 01 January 2024, data from June-July 2022 to May 2023 were used. All models then used the same length of estimation window. Recursive (iterative) multi-step forecasts were run, with all dynamic models using previous load forecasts as input for forecast horizons longer than one day. Regarding the impact of meteorological factors such as temperature and humidity on load forecasting, two approaches were compared. The first one was to estimate the lower bound of the forecast error for the model with this variable, assuming that the temperature can be accurately predicted. The second one used a simple method to predict the temperature and evaluated how the load forecasting performance changes over different time horizons when the predicted temperature is used instead of the actual temperature.

[0076] In order to improve the prediction accuracy of the model, it is necessary to conduct a preliminary analysis of the correlation between various influencing factors and load, and eliminate factors with small correlation to reduce noise. To this end, the correlation between different influencing factors and power load is quantified. Given that some characteristics are nonlinear influencing factors, the Copula function can be used to analyze the nonlinear relationship between variables. The Copula function characterizes the correlation between nonlinear and asymmetric variables from a probability perspective, which is suitable for random variables of any distribution and has advantages in analyzing the correlation between loads with volatility and randomness and their influencing factors.

[0077] After determining the marginal probability distribution of load and influencing factors, the Copula function is used to obtain the joint probability distribution function of random variables, and it is applied to the rank correlation coefficient to obtain the correlation measure. The Gaussian Copula function suitable for multiple variables is used to obtain that the closer the rank correlation coefficient is to 1, the stronger the correlation between the two variables. The factors with strong correlation are: maximum temperature, minimum temperature, average temperature, humidity, and holiday type. Therefore, it is necessary to introduce the maximum temperature, minimum temperature, average temperature, humidity, and holiday type as input feature quantities in the prediction model.

[0078] After inputting the sample load data into the FNN model, the load at each moment of a typical day is predicted. The prediction results and error analysis are as follows.

[0079] like Figure 1 and Figure 2 .

[0080] Analysis shows that in terms of relative error, the average, maximum and minimum values ​​are 4.51%, 7.59% and 0.17% respectively. The error is within the allowable range, and after adding relevant influencing factors, the accuracy is significantly improved.

[0081] In the analysis of influencing factors and load forecasting, it was found that incorporating multiple factors helps to improve the prediction accuracy of the model. The error of a single feature is large when the time scale to be predicted is large. This is because there is an interaction between the input features, which causes interference in the model when identifying a feature with a large influencing factor, thereby affecting the prediction accuracy. Therefore, multiple influencing factors are considered, and sub-models are trained through category labeling based on different feature properties to identify the time period to be predicted. Different weights are assigned to the sub-models in different time periods to finally obtain a combined model.

[0082] Based on the FNN prediction model, the historical load data is taken into account, and the load influencing factors, such as weather conditions and holidays, are comprehensively considered to predict the load under different types of days.

[0083] like Figure 3 ,From the load forecast results for holidays, we can see that the absolute percentage error of the forecast error at 17 moments is less than 7%, the minimum absolute percentage error is 1.38%, the maximum absolute percentage error is 10.47%, and the average absolute percentage error is 4.92%, indicating that the forecast effect is relatively ideal. Figure 4 The load forecast results for working days show that there are 18 moments with an absolute percentage error of less than 7%, the minimum absolute percentage error is 1.47%, the maximum absolute percentage error is 9.41%, and the average absolute percentage error is 4.46%. The forecast results are relatively ideal.

[0084] By observing the model fitting effect on the training set, it is found that the model with multiple features has a better fitting condition on weekdays, but a poorer fitting condition on holidays. For this reason, the feature input method can be adjusted, and the model with the "holiday" feature can be trained separately. It is observed that the prediction accuracy is significantly improved after adding the refined "holiday type" model.

[0085] Figure 5 This is a comparison chart of the typical daily load forecast curve after considering real-time meteorological factors, the actual load curve and the FFN load forecast curve. It can be seen from the figure that the daily load curve after meteorological load correction has a more obvious error reduction compared with the FFN load forecast curve. For specific data, see Figure 6 According to the error analysis of the forecast results, the maximum relative error after correction is 6.27%, the minimum relative error is 1.02%, and the average absolute relative error is 3.89%, which is lower than the maximum allowable average error standard (5%). Therefore, the load forecasting model after meteorological factors correction has higher prediction accuracy than the FFN load forecasting model.

[0086] Using parameters consistent with the single-characteristic variable load forecasting model, all load data from 2022 to 2024 that have been cleaned and processed and the screened influencing factors are simultaneously input into the established FNN model, with the training set and test set divided in a ratio of 8:2. As the training progresses, the Loss of both the training set and the test set decreases significantly and eventually flattens out, indicating that the model has reached a convergence state.

[0087] The present invention establishes a medium-term load forecasting model that takes into account the influencing factors of meteorology and holidays, integrates the load characteristics caused by meteorology, holidays and other factors into the forecasting model, and realizes the organic combination of multiple influencing factors and load variation laws. In view of the limitations of a single load forecasting method, in actual operation, the forecasting algorithm performs combined forecasting based on actual conditions, and complements the advantages of multiple algorithms. In terms of load influencing factors, two aspects of meteorological factors are considered: humidity and temperature. The present invention takes into account the impact of multiple factors on the load and improves the accuracy of load forecasting.

[0088] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A medium-term load forecasting method based on feedforward neural network and time series, characterized in that: The steps include: Step 1: Establish an inbound and outbound power base load prediction model; Step 2, establishing a prediction and correction model for the impact of temperature accumulation effect on load; Step 3: Establish a prediction and correction model for the impact of humidity changes on load; Step 4, establish a meteorological sensitive load prediction model after considering the relationship between humidity, temperature and load changes; Step 5, establishing a total load forecasting model; Step 6: Use sample matching data to evaluate the accuracy of the model.

2. A medium-term load forecasting method based on a feedforward neural network and time series according to claim 1, characterized in that: Step 1: Establishing an in-and-out power base load forecasting model is specifically as follows: using a multi-layer feedforward neural network method to establish a medium-term power load forecasting model.

3. A medium-term load forecasting method based on feedforward neural network and time series according to claim 2, characterized in that: Select a multi-layer feedforward neural network to build a multi-layer artificial neural network model, including input layer, hidden layer and output layer. The specific model is: Where: y t represents the output variable at time t, X represents the vector including the lagged output variable and the external variable input, n represents the number of input variables, and h represents the number of neurons in the hidden layer; f() and g() are S-type nonlinear activation functions; in addition, v j and w ij are weight parameters, representing the connection strength between hidden nodes and output and between input nodes and hidden nodes, b and b j is the bias parameter for the intercept in the linear model.

4. A medium-term load forecasting method based on feedforward neural network and time series according to claim 3, characterized in that: The model prediction accuracy and performance were evaluated by calculating the root mean square error and mean absolute percentage error; RMSE is the root mean square error, N is the number of data observations in the window for prediction accuracy comparison, and l t is the load predicted at time t, and MAPE is the mean absolute percentage error.

5. The medium-term load forecasting method based on feedforward neural network and time series according to claim 1 is characterized in that: Step 2: The prediction and correction model for the impact of temperature accumulation on load is as follows: Where: T is the corrected temperature value; i represents the i-th hour before the predicted time, T i represents the temperature value one hour before the predicted time; T jd represents the temperature value j days before the forecast date; m i 、n j They correspond to the temperature sensitive load change and human perception hysteresis correction coefficient respectively.

6. A medium-term load forecasting method based on feedforward neural network and time series according to claim 1, characterized in that: Step 3: The prediction and correction model for the impact of humidity changes on load is as follows: Where: H is the humidity value of the forecast period after correction; H0 is the humidity value of the forecast period before correction; H i is the humidity value of the i-th period before the prediction; k i is the weight parameter; P is taken as 1, which means that only the humidity values ​​in the first two periods of the forecast are taken into account.

7. A medium-term load forecasting method based on feedforward neural network and time series according to claim 1, characterized in that: Step 4: After considering the relationship between humidity, temperature and load changes, the meteorological sensitive load prediction model is as follows: Where: L w is the meteorological sensitive load; C is a constant; H is the relative humidity value obtained after correlation correction; T h is the maximum temperature during the forecast period; T a is the average temperature during the prediction period, and k1, k2, k3, k4, and k5 are weight parameters.

8. The medium-term load forecasting method based on feedforward neural network and time series according to claim 1 is characterized in that: Step 5: The total load forecasting model is as follows: L=L i +L w Where: L is the total load value, L i is the basic load value, L w is the corrected meteorological sensitive load value.

9. The medium-term load forecasting method based on feedforward neural network and time series according to claim 1 is characterized in that: In step 6, the Copula function is used to analyze the nonlinear relationship between variables, and the influencing factor analysis is evaluated on the model correlation analysis to obtain the factors with strong correlation.

10. A medium-term load forecasting method based on feedforward neural network and time series according to claim 9, characterized in that: The factors with strong correlation are: maximum temperature, minimum temperature, average temperature, humidity, and holiday type.

Images