CN112865093B - Combined prediction method for short-time power load - Google Patents

Abstract

The invention relates to a short-time power load prediction method, which is characterized in that according to the correlation between meteorological factors and power loads, total power loads are split into basic loads and meteorological sensitive loads, the basic loads are subjected to load prediction by adopting an autoregressive time sequence method, the meteorological sensitive loads are subjected to load prediction by adopting a support vector machine algorithm, and the two predicted loads are overlapped to realize the prediction of the total power loads. Compared with the traditional short-time power load prediction method, the method has the advantages that the power load is split into the base load and the weather-sensitive load, the prediction model is respectively built aiming at different load types, the accuracy of the power load prediction model is improved, and the annual applicability of the prediction method is improved.

Description

A short-term power load combined forecasting method

Technical field

The present invention relates to the technical field of power load forecasting, and in particular to a short-term power load combined forecasting method.

Background technique

Electricity is the basic energy related to the national economy and people's livelihood. A safe and reliable power system is an important guarantee for national economic growth and social development. The electric power system is different from general industrial industries in that the electric energy produced cannot be stored on a large scale. This also requires the production and consumption of electric energy to achieve a real-time balance, that is, to maintain on-demand supply at all times. The essence of load forecasting is to predict the needs of users in the power system. Accurate load forecasting is closely related to all levels of the power system and plays a decisive role in the normal operation of the power system. The forecast of demand-side power load is the benchmark for power generation companies, power transmission and distribution departments, dispatch centers and other departments in power planning, formulating power generation plans, arranging generator unit maintenance plans, planning cross-regional power transactions, etc. It is an important step to ensure the security of power supply and The basis of reliability and economy.

Electric power load forecasting is divided into the following four categories according to the length of the future prediction time: 1. Ultra-short-term load forecasting, which is to predict the load within one hour in the future, is mainly used for emergency dispatch and demand-side response adjustment in the power market; 2. Short-term load forecasting, whose forecast target is the power load from one to several days in the future, mainly serves to optimize the power generation plan on the power supply side, efficiently arrange the operation of generating units, facilitate the reasonable arrangement of dispatch plans on the power transmission and distribution side, and at the same time provide information for short-term power price fluctuations. Basis; 3. Medium-term load forecasting refers to predicting the power load from one month to one year in the future, which facilitates the power supply side to arrange longer-period power generation plans and unit maintenance and other plans; 4. Long-term load forecasting refers to 3 to 5 years in advance. The annual load forecast is mainly used for long-term planning of the power system to facilitate planning of power supply capacity expansion, power grid construction and other projects.

Previous short-term power load forecasting models mainly used data-driven forecasting models, ignoring the physical implications such as the composition and changing characteristics of the power load itself. The establishment of data-driven models depends on the quality of data samples. In actual situations, the number of data samples is often limited, which results in low generalization of existing model algorithms. At the same time, existing short-term power load forecasting models do not comprehensively consider the influencing factors of power load, and only consider the impact of temperature on power load. Therefore, it is necessary to develop a short-term power load forecasting model based on the physical characteristics of power load composition and comprehensively considering the impact of various weather factors on power load to improve the generalization and prediction accuracy of the algorithm.

Contents of the invention

The technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology and propose a short-term power load combined prediction method.

In order to solve the technical problem, the solution of the present invention is:

A short-term power load combined forecasting method. The specific process is shown in Figure 1 and includes the following steps:

Step 1, determine the seasonal attributes of the day;

Step 2, load prediction;

In step 2, the load prediction step specifically includes:

Step 2.1, base load forecast;

Step 2.2, weather-sensitive load forecasting.

In step 1, the day's seasonal attribute determination uses the day's season attribute determination model to determine the day's season in the future day. The specific process is shown in Figure 2, which is determined by the mode weighted average of the hourly seasonal attributes in the future day. Determine the seasonal attributes S _D of the forecast day:

Among them, S _T is the seasonal attribute at a certain time; p is the weighted coefficient of the seasonal attribute at different time periods; n is the number of seasonal attribute modes at different times in the future day; i represents each moment corresponding to the seasonal attribute mode.

The seasonal attributes of the moment are calculated through the hourly seasonal attribute determination method. The hourly seasonal attribute determination adopts a decision tree model. The input parameters are the sliding average effective temperature (mET, ℃) in the first three hours and the sliding average air enthalpy in the first three hours. value (mEnth, kJ/kg), the sliding average moisture content in the first three hours (md, g/kg) and the sliding average dew point temperature in the first three hours (mTs, ℃). The seasonal division S _T at a certain moment is obtained through the decision rule. The seasonal division is defined as a discrete attribute, and the corresponding physical meaning is as shown in the following formula:

The determination rules for the seasonal division are as shown in Figure 3. The root node uses the effective temperature as the root node attribute according to the information gain of each attribute. When the sliding average effective temperature at the current moment is less than 9.8°C, the data returns to the left branch, towards the left branch. It is divided into heating season and transition season. According to research on human thermal comfort, when the effective temperature is less than 9°C, the human body will feel cold. The second-level node examines the air enthalpy value. When the left branch reaches the leaf node when the enthalpy value is less than 43kJ/kg, it can be determined that heating is required at the current moment and it is the heating season. On the contrary, the unique attribute category is obtained based on whether the moisture content of the third-level node is greater than or less than 10g/kg, thereby dividing the heating season and the transition season. The effective temperature of the root node data is greater than 9.8°C, and the data flows to the right branch, thus finally dividing the cooling season and the transition season. The largest information gain of the secondary node on the right is also the enthalpy value. When the enthalpy value is less than 20.8kJ/kg, it can be judged as a transition season. On the contrary, it continues downward to the third season point. When the effective temperature is greater than 22°C, it is judged to be the cooling season. When the effective temperature is less than 22°C, the data subset of the fourth layer node is divided into transition season and cooling season based on whether the moisture content is greater than 15g/kg.

In the time-by-hour seasonal attribute determination decision tree model, in order to reduce the over-fitting phenomenon of the decision tree, n-fold cross-validation is used during model training. The training samples are randomly divided into n subsets, of which n-1 subsets are used for training, and a separate sample is used as data to verify the model. The cross-repetition is repeated n times, and the results of n times are integrated to optimize the model.

The weighted coefficients of seasonal attributes in different periods are determined based on the possibility of people turning on the air conditioner at different periods. According to the daily peak and valley changes, the day is divided into the nighttime low period (0:00-7:00) and the daytime peak period (8:00-7:00). :00-18:00) and evening peak hours (19:00-23:00). The weighting coefficients of seasonal attributes in different periods are:

In step 2.1, the basic load is a load that is greatly affected by daily life schedules and has little correlation with meteorological changes. It is divided into working day basic load and holiday basic load according to different types of days. The weighted average daily load at each time of the historical data corresponding to the day type in the transition season is used as the basic load part of the power load. Using the seasonal attribute determination model, the daily basic load Q _{F_D} calculation formula is:

Among them, it is considered that when the daily seasonal attribute S _D = 0, it is the transition season, and the daily average load in the transition season is the daily basic load; when 0 < | S _D | < 0.5, it is the incomplete transition season, and the daily basic load is the daily basic load. The average value of the base load at each time, the base load Q _{F_Ti} at each time on the above-mentioned day is:

Among them, T _i represents a certain moment; S _Ti is the seasonal attribute value corresponding to a certain moment; n is several moments used for calculation before _Ti moment; is the basic load at time T _im .

In step 2.1, the base load prediction uses a base load prediction model to predict the base load through the time series method, and the input data is the base load calculated by the aforementioned method for the same type of days adjacent to the prediction day. For stationary time series data, the main prediction models are MA(q), AR(p) and ARMA(p, q); for non-stationary time series data, the trend change or seasonal change of the data is removed by difference, and the A new stationary time series sample is used to establish an ARMA prediction model. The main prediction model is ARIMA. When determining the model order, the Akaike information criterion (AIC) is used to estimate p and q, and p and q under the minimum AIC are obtained as model parameters.

The weather-sensitive load refers to the seasonal attribute of the day within the same day type grouping. If it is the heating season or the cooling season, the base load is subtracted from the day's load, and the excess is the weather-sensitive load. Meteorologically sensitive loads are predicted using the support vector machine model (SVR), and the grid search method is used for model parameter optimization. The optimized parameters in this model include the penalty factor C and the kernel function parameter γ. The formula is as follows:

Among them, σ is the kernel width.

In step 2.2, the meteorological-sensitive load prediction uses the input parameters of the meteorological-sensitive load prediction model that change with the seasonal type of the day on the prediction day. When the seasonal type of the day is the heating season (SD=1), the input parameters of the model are The variables are shown in the following formula:

S _D =1:

Among them, T _i is the prediction time; mET is the moving average effective temperature of the four points before the prediction time; is the weather-sensitive load at time T _ij . The input space is six dimensions.

When the daily season type is the cooling season (S _D =-1), the variables entered into the model are as follows:

Among them, mT is the moving average temperature of the four points before the prediction time; is the moving average relative humidity at the four points before the prediction time; mν is the moving average wind speed at the four points before the prediction time. The input space is eight-dimensional.

The short-term power load combined forecasting method can achieve 15-minute power load forecast for the next day, and the total predicted load at each moment is:

Q _{All_T} = Q _{F_T} + Q _{W_T}

Among them, Q _{All_T} is the forecast value of the power load at a certain time in the future day; Q _{F_T} is the basic load forecast value of the electric power load at a certain time in the future day; Q _{W_T} is the weather-sensitive load forecast value of the electric power load at a certain time in the future day.

Description of drawings

Figure 1 is a flow chart of the short-term power load combined forecasting method.

Figure 2 is a flow chart of the method for determining the seasonal attributes of the day.

Figure 3 shows the season division determination rules at a certain moment.

Figure 4 Model confusion matrix.

Figure 5 is a time series comparison diagram of the data prediction value and the target value in the embodiment.

Figure 6 shows the optimization results of the SVR model parameter penalty factor C and kernel function distribution width γ.

Figure 7 shows the fitting degree and deviation of the model on the training set.

Figure 8 shows the fitting degree and deviation of the model on the test set.

Detailed ways

Figure 1 is a flow chart of the short-term power load combined prediction method of the present invention.

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments (taking the power load every 15 minutes in Shanghai in 2014 as an example for training and prediction, the data comes from the Shanghai Municipal Economic and Information Commission data platform), specifically including the following: step:

Step 1: Train the seasonal attribute determination model for the day. Based on the background knowledge of building energy consumption, the heating season in Shanghai is determined to be January, the cooling season is determined to be July, and the transition season will appear in April and October. Therefore, the seasonal attribute of January data is assigned a value of 1, and the seasonal attribute of July data is assigned a value of -1. According to the human body comfort index, when the effective temperature ET falls between 15-23°C, the human body feels more comfortable, so it can be considered that April and October The effective temperature period within this range is the transition season. If the effective temperature in April and October is not within this range, it is temporarily considered that the seasonal attributes cannot be determined. The data of this period can be classified through subsequent decision tree rule mining of the training data.

The input training data is hourly data in January and July, as well as data with effective temperatures falling between 15-23°C in April and October. The seasonal discrimination rules generated by the decision tree are shown in Figure 3. After training, the classification accuracy of the generated decision tree is 92.5%. The model confusion matrix is shown in Figure 4. The model's overall classification error for the transition season is 10%, among which the classification error for the transition season and heating season is 7%. This is mainly due to the transition season weather parameters and The weather parameters in the heating season overlap more, and the left branch of the decision tree has fewer levels. However, in order to prevent data overfitting and for hour-by-hour seasonal judgment, the accuracy of the model has met the requirements.

Step 2: Train the base load forecast model. Take the data of five consecutive working days from March 17 to March 21 as an example to conduct training modeling, and predict the load on March 24 to analyze the prediction performance of the model, where the interval dates are holidays. After judging the stationarity of the data, identifying the model and determining the order, the form of the model was determined to be ARIMA (5, 2, 3). The coefficients obtained by the least squares method are shown in Table 1:

Table 1 Determination of working day sample prediction model parameters

After calculating the specific form of the forecast model, the 96-point data on March 24 was forecast and compared with the actual base load on that day in order to evaluate the performance of the forecast model. The model evaluation indicators for the test set are shown in Table 2. Overall, the time series method is used to predict the base load in the transition season with high accuracy, the relative error is less than 2%, and the model has a high degree of fit. It shows that this method has better performance in predicting base load with stable diurnal variation.

Table 2 Prediction model evaluation table for working day sample test set

Figure 5 shows the time series comparison chart of the predicted value and the target value of the 96-point data on March 24. The predicted value shown in the figure below is highly consistent with the target value, and can better predict the changing characteristics of the base load.

Step 3: Train the weather-sensitive load forecasting model to predict the total power load. Taking the heating season prediction model as an example, in order to ensure a sufficient number of training samples, 960 data items for ten days from January 7 to January 16 were selected as the training set, and the data on January 17 and January 20 were used as the test set. The separated dates are excluded holidays. Figure 6 shows the optimization results of the SVR model parameter penalty factor C and kernel function distribution width γ. The minimum standardized mean square error (MSE) of the model is targeted through the grid search method. The obtained C and γ are the inputs. optimal parameters under the sample. For the above training samples, the optimal equation parameters are C=2.828 and γ=5.67.

Table 3 gives a summary of the performance parameters of the weather-sensitive load forecasting model based on the training set and prediction set. The 5-fold cross-validation method is used when fitting the model on the training set. The trained model has better performance. First of all, ^R2 is higher and has a higher degree of fitting. Secondly, the RMSE of the predicted value and the target value is small, indicating that the residual distribution is relatively concentrated, and the MAE is 170.8, and the average relative error is 12%. Relatively speaking, the prediction performance of the model on the test set has declined, but the comparison of the average relative error, combined with the small number of input parameter dimensions and the large enough training sample, shows that the model has good generalization and eliminates the problem of excessive fitting situation. The RMSE and MAE in the test set are larger, mainly because the fluctuations of meteorological sensitive loads are more random and the periodic variability is not strong. However, comparing the data of the predicted meteorological sensitive load and base load combined with the total power load sequence, the MAPE is significantly reduced, which greatly optimizes the relative error in the overall forecast set.

Table 3 Performance summary of meteorologically sensitive load forecasting models

Figures 7 and 8 show the fitting degree and deviation of the model on the training set and test set respectively. Overall, the model can better predict the changing trend of the power load.

Claims (4)

1. A short-time power load combined prediction method is characterized in that the flow comprises the following steps:

step 1, judging the season attribute of the current day;

in step 1, the current day season attribute determination is performed by using a current day season attribute determination model for determining the current day season of the future day, and determining the season attribute S of the predicted day by a mode weighted average of the time-by-time season attributes of the future day _D ：

Wherein S is _T Is a seasonal attribute at a certain moment; p is the season attribute weighting coefficient of different time periods; n is the number of season attribute modes at different times of the future day; i represents each moment corresponding to the season attribute mode;

step 2, load prediction; in step 2, the load prediction step specifically includes:

step 2.1, predicting a base load;

in the step 2.1, the base load is a load which is greatly influenced by life work and rest and has smaller correlation with weather change, and is divided into a workday base load and a holiday base load according to different day types; adopting daily load of weighted average of each moment of historical data of corresponding daily type in transition season as a basic load part of the power load; adopting the seasonal attribute determination model, the daily base load Q _{F_D} The calculation formula is as follows:

wherein, consider the day-season attribute S _D When=0, the average daily load in the transition season is the base daily load; when 0 < |S _D When the level is less than 0.5, the time of day base load is the average value of the base load at each time of the day, and the base load Q at each time of the day is the incomplete transition season _{F_Ti} The method comprises the following steps:

wherein T is _i Indicating a certain moment; s is S _Ti The season attribute value corresponding to a certain moment; n is T _i A plurality of moments before the moment for calculating;is T _i-m A moment base load;

and 2.2, weather-sensitive load prediction.

2. The short-term power load combined prediction method according to claim 1, wherein in step 1,

the seasonal attribute at the moment is calculated by a time-by-time seasonal attribute judging method, a decision tree model is adopted for judging the time-by-time seasonal attribute, and input parameters are a first three-hour moving average effective temperature mET (DEG C), a first three-hour moving average air enthalpy value mEnth (kJ/kg), a first three-hour moving average moisture content md (g/kg) and a first three-hour moving average dew point temperature mTs (DEG C); obtaining seasonal division S at a certain moment through judging rules _T Seasonal divisions are defined as discrete attributes, corresponding physical meanings are shown in the following formula:

according to the season division, the root node takes the effective temperature as the root node attribute according to the information gain degree of each attribute, when the current time sliding average effective temperature is less than 9.8 ℃, the data is led to the left branch, and the heat supply season and the transition season are divided downwards; according to the related study of the thermal comfort of the human body, when the effective temperature is less than 9 ℃, the human body can generate cold feeling; the second-level node examines the enthalpy value of the air, and when the enthalpy value is less than 43kJ/kg, the left branch reaches the leaf node, so that the heating is required at the current moment and the heating season is judged; on the contrary, the unique attribute category is obtained through whether the moisture content of the third-stage node is more than or less than 10g/kg, so that a heating season and a transition season are divided; the effective temperature of the root node data is more than 9.8 ℃, and the data branches to the right side to flow, so that a refrigerating season and a transition season are finally divided; the maximum gain of the right-side secondary node information is also an enthalpy value, and when the enthalpy value is smaller than 20.8kJ/kg, the transition season can be judged; otherwise, continuing to downwards reach the third-stage node, and judging that the cooling season is performed when the effective temperature is higher than 22 ℃; dividing the subset of data of the fourth level node into a transition season and a refrigeration season by whether the moisture content is greater than 15g/kg when the effective temperature is less than 22 ℃;

in the time-by-time seasonal attribute decision tree model, n-fold cross validation is adopted in model training in order to reduce the overfitting phenomenon of the decision tree; randomly dividing the training samples into n subsets, wherein n-1 subsets are used for training, another single sample is used as data of a verification model, the training samples are alternately repeated for n times, and the result optimization model is synthesized for n times;

the season attribute weighting coefficients of different time periods are determined according to the possibility that people turn on air conditioner in different time periods, and the time periods are divided into night valley time periods 0:00-7:00, daytime peak time periods 8:00-18:00 and evening peak time periods 19:00-23:00 according to daily peak-valley changes; the season attribute weighting coefficients of different time periods are:

3. the short-term power load combined prediction method according to claim 2, characterized in that,

in step 2.1, the base load is predicted by adopting a base load prediction model through a time sequence method, and input data is the base load of the same type of day adjacent to the predicted day calculated by the method; for stationary time series data, the prediction model is MA (q), AR (p) and ARMA (p, q); for non-stable time sequence data, establishing an ARMA prediction model for the obtained new stable time sequence sample through differential removal of trend change or seasonal change of the data; when determining the model order, estimating p and q by adopting a red pool information criterion AIC (Akaike information criterion) to obtain p and q under the minimum AIC as model parameters;

the weather-sensitive load refers to the weather-sensitive load, the weather-sensitive load is divided into groups of the same day type, the current day season attribute is judged, if the weather-sensitive load is a heating season or a cooling season, the current day load is used for subtracting the base load part, and the excess part is the weather-sensitive load; the meteorological sensitive load is predicted by adopting a support vector machine (SVR) model, the model parameter optimization adopts a grid search method, and the optimized parameters in the model comprise a punishment factor C and a kernel function parameter gamma, and the formula is as follows:

wherein σ is the kernel width.

4. The short-term power load combined prediction method according to claim 3, characterized in that,

in step 2.2, the input parameters of the weather-sensitive load prediction model are changed along with the different types of the seasons of the day of the prediction, and when the types of the seasons of the day are heat supply seasons, the variables of the input model are shown in the following formula:

wherein T is _i Is the predicted time; mET is the effective temperature of four points of moving average before the predicted time;is T _i-j Weather-sensitive load at time;

when the type of the day season is the cold season, the variables of the input model are as follows:

wherein mT is the four-point moving average temperature before the predicted time;the four points before the predicted moment are the moving average relative humidity; mν is the four-point sliding average wind speed before the predicted time;

the short-time power load combined prediction method can be used for predicting the power load 15 minutes a day in the future, and the predicted total load at each moment is as follows:

Q _{All_T} ＝Q _{F_T} +Q _{W_T}

wherein Q is _{All_T} A predicted value of the power load at a certain time of day in the future; q (Q) _{F_T} A base load predicted value for the power load at a time of day in the future; q (Q) _{W_T} And weather-sensitive load forecast values for the power load at a certain time of the future day.