CN107730031A - A kind of ultra-short term peak load forecasting method and its system - Google Patents

Abstract

The present invention relates to a method and system for ultra-short-term peak load forecasting, which collects weather forecast information; performs ultra-short-term load forecasting based on the established peak load forecast model and the collected weather forecast information; the peak load forecast model includes: Load index, meteorological factors affecting peak load and correlation of each meteorological factor; based on the peak load and meteorological factors affecting peak load. The actual calculation example provided by the invention proves that the method is effective for peak peak prediction and improves the prediction accuracy of ultra-short-term load.

Description

A method and system for ultra-short-term peak load forecasting

technical field

The invention relates to a forecasting method and system in the field of electric power system dispatching automation, in particular to an ultra-short-term peak load forecasting method and a system thereof.

Background technique

With the economic growth and the improvement of people's living standards, the power load is increasing year by year. Ultra-short-term load forecasting uses the latest load information to make real-time forecasts of power loads within 5 minutes to 4 hours in the future. Ultra-short-term load forecasting can track changes in power system load online, and is the basis for dynamic power grid security monitoring and automatic power generation control. Accurate and fast ultra-short-term load forecasting plays an important supporting role in ensuring the safety and economy of power grid operation.

At present, the main methods of ultra-short-term load forecasting include extrapolation method, support vector machine, neural network and data mining. Because these prediction models believe that the weather changes little when the time is short, and the accuracy of the weather forecast is not high, it is easy to cause error superposition, so most of them do not consider the impact of meteorological factors on the load. Therefore, the prediction accuracy of the above methods is relatively high in the period of stable load, but the prediction error is relatively large in the load periods that are particularly sensitive to meteorological changes, such as peak and trough loads.

The difficulty of ultra-short-term load forecasting still lies in peak load forecasting. Through the research on the typical summer load of a certain province, it is found that due to the sudden change of the weather in the province due to many storms in summer, the city's cooling load suddenly decreases, resulting in a decrease in the total load. According to conventional forecasting methods, it is obviously difficult to meet the practical requirements of ultra-short-term forecasting without considering sudden changes in weather or only considering daily characteristic meteorological elements, such as daily maximum temperature, daily minimum temperature, and weather types. Therefore, accurate real-time weather forecast becomes a key factor to improve the accuracy of load forecasting.

Numerical weather prediction is a more refined weather forecast in time and space, and it is mainly used in wind power prediction in power systems. It provides a wealth of meteorological information, including various information such as air pressure, temperature, humidity, wind, cloud, and precipitation per tens of square kilometers on the spatial scale; on the temporal scale, the prediction results are as fine as hours or less. Numerical weather prediction provides real-time and accurate meteorological data support for power grid ultra-short-term load forecasting.

Because peak load is very sensitive to weather changes, peak load forecasting has always been a difficult problem in power systems due to its randomness and complexity.

Contents of the invention

In order to solve the deficiencies in the above-mentioned prior art, the object of the present invention is to provide a super-short-term peak load forecasting method and its system, which is effective for peak-peak forecasting and improves the forecasting accuracy of super-short-term load.

The object of the present invention is to adopt following technical scheme to realize:

The present invention provides an ultra-short-term peak load forecasting method, the improvement of which is:

Collect weather forecast information;

Perform ultra-short-term load forecasting based on the established peak load forecasting model and the collected meteorological forecast information;

The peak load prediction model includes: peak load index, meteorological factors affecting peak load and the correlation of each meteorological factor; based on the peak load and meteorological factors affecting peak load.

Further, the peak load forecasting model includes:

Determine the peak load index;

Based on the historical meteorological information, analyze the peak load and the meteorological information to determine the correlation of each meteorological factor affecting the peak load;

Based on the peak load and the meteorological factors affecting the peak load, a peak load prediction model is constructed in combination with the L-M neural network.

Further, the determination of the peak load index includes:

Determine the peak load based on pre-acquired data and analyze the peak load to extract peak load indicators;

The peak load indicators include:

1) Peak and valley load values:

The peak load value P _peak is the largest load value in a day; the valley load value P _valley is the smallest load value in a day;

2) Peak moment:

The peak moment is the moment when the peak load peak occurs;

3) Peak-to-valley difference:

The expression of the peak-to-valley difference is as follows:

P _difference = P _peak – P _valley (1)

Among them, P _difference is the peak-to-valley difference;

4) The steep rise rate of high load, including:

The system load P _h satisfies the relationship:

P _h ≥ P _peak – P _difference /4 (2)

At this time, the load level is at a high level;

The high load slope at time t is:

In the formula: s is the high load slope, d is the high load time interval, P _t represents the high load value at time t, P _t represents the high load value at time t-1; s _t is the high load slope at time t;

The load slope of the high-level steep rise section must reach the steep rise level, and the steep rise level is measured by the following formula:

s _steep ≥0.75s _max (4)

In the formula: s _max is the slope of the steepest period;

5) The starting time of high load steep rise:

The expression is as follows:

T _steep ＝min(T _steep1 ,T _steep2 ,…,T _steepn ) (5)

Among them, s _steep is the level of steep rise; T _steep1 , T _steep2 ,...,T _steepn are respectively: the first, second,..., the beginning moment of the nth high steep rise segment; T _steep is the sample curve At the beginning of the high-level steep rise, the ratio of the load increment to the time increment of the high-level steep-rise section that satisfies formula (4) is defined as the high-level steep rise rate;

6) Peak duration:

Peak duration is the time span of load formation that reaches the high load level.

Further, based on the historical weather information, analyzing the peak load and weather information to determine the correlation of meteorological factors affecting the peak load, including:

Determine relevant meteorological factors according to the historical meteorological information;

Analyze the correlation between peak load and meteorological factors;

Based on the robust regression model, the correlation of the meteorological factors affecting the peak load is verified.

Further, the meteorological factors include: instantaneous wind speed, instantaneous temperature, instantaneous air pressure, maximum air pressure, maximum wind direction, maximum wind direction, rainfall, maximum wind speed, instantaneous wind direction, instantaneous humidity, minimum humidity, maximum temperature, minimum temperature, minimum Barometric pressure, maximum wind speed information.

Further, the analysis of the correlation between peak load and meteorological factors includes: calculating the correlation coefficient of each meteorological factor based on the following formula:

In the formula: R is the correlation coefficient; cov(X,Y) is the covariance of X and Y; are the mean square errors of X and Y, respectively;

X is: historical load data; Y is: historical temperature data.

Further, the robust regression model includes:

Construct a robust regression model calculation formula, as follows:

Y=Xβ+ε (7)

Among them: Y=(y _i ) _m×1 is the historical load data at a certain time in m days; X=(x _ij ) _m×n is the historical temperature data at a certain time in m days; β=(β _j ) _n×1 is estimated Unknown parameter vector, ε=(ε _i ) _m×1 is an unobservable random error vector;

Based on the following formula, estimate the parameters of β=(β _j ) _n×1 in the robust regression model:

Among them: D(β) represents the objective function in the robust regression model, y _i represents the historical load data at a certain time on the i-th day, i=1,...,m, x _ij represents the temperature data at a certain time on the i-th day, and m represents total number of days;

Finding the Best Estimated Parameters The key is to determine the weighting function matrix; let the m-order full-rank diagonal matrix W=(w _i ) _m×m be the weighting function matrix, and obtain the optimal estimated parameters by solving Solving formula:

in, r _i =y _i -x _(i,j) β _j is the residual item; X ^T is the transposed data of the historical temperature data X at a certain moment in m days.

Further, the peak load prediction model is constructed based on the peak load and the meteorological factors affecting the peak load in combination with the L-M neural network, including:

Based on the peak load and the meteorological factors affecting the peak load, select the training samples of the L-M neural network;

determining a data type vector for the composite data sample;

Build a peak load forecasting model based on training samples.

Further, the selection of training samples of the L-M neural network based on the peak load and the meteorological factors affecting the peak load includes: selecting system load samples and corresponding comprehensive data samples of k working days or holidays before the forecast date ;

The integrated data type vector of the integrated data sample is:

D _k ＝(D _k1 D _k2 D _k3 D _k4 D _k5 ) ^T (16)

Among them: k represents the number of days k=1,2,...,p, p is the number of days selected, D _k1 is the daily maximum temperature on the kth day, D _k2 is the daily minimum temperature on the kth day, and D _k3 is the temperature on the kth day Weather conditions, D _k4 is the average humidity of the k-th day, and D _k5 is the day type of the k-th day;

Build a peak load forecasting model based on training samples, including:

Through the C-means fuzzy clustering method, the vectors belonging to the same class are obtained through iterative optimization calculation from the comprehensive data type vectors of the comprehensive data samples. In this class, the peak load sample curves of the days corresponding to the vectors are selected and recorded as Y respectively. _p1 (t), Y _p2 (t), ..., Y _pm (t);

The peak load forecasting model is determined according to the peak load sample curve, and the expression is as follows:

M _p = {Y _p1 (t), Y _p2 (t),..., Y _pm (t)} (17)

Among them: Y _p1 (t), Y _p2 (t), ..., Y _pm (t) are the first, second, ..., m peak load sample curves respectively; m is the peak load The number of sample curves.

Further, ultra-short-term load forecasting is performed based on the peak load forecasting model combined with near weather forecasting, including:

Select the vectors that are input to the peak load forecasting model;

Calculate the peak load forecast error until the iteration condition is met;

Output peak load forecast results.

Further, the selection of the vector input to the peak load forecasting model includes:

Taking the time threshold as each interval period of the peak load, the prediction object is the set point load of a day; for a single training, the load at the predicted time of m peak samples is used as the load input vector of the neural network;

According to the correlation analysis between the peak load and meteorological factors, the temperature t _k at the same time, the rainfall r _k at the same time, the maximum daily temperature t _max , and the minimum daily temperature t _min of m peak load samples are selected as the peak load forecast The meteorological factors input to the model.

Further, the time threshold is selected as 5 minutes; the set point load is selected as 288 points.

Further, the iteration condition is expressed by the following formula:

Among them: Y(t _i ) is the output value of the peak load forecasting model, A(t _i ) is the ideal value, W is the optimal value that minimizes the error between the output value of the peak load forecasting model and the ideal value, and ε is the forecasting accuracy threshold ; i represents the i-th day.

Further, the peak load prediction results include:

Load Forecast Power Estimates

For n peak load samples whose prediction category is consistent with the forecast date and belong to the same model, calculate the daily high and steep start time vector T _steep(1×n) = (T _steep1 T _steep2 …T _steepn ) ^T , and calculate the average Start time T _Avsteep ;

During peak hours [T _AVsteep -T _2H , T _AVsteep +T _2F ], where T _2H takes a value between 0.5 and 1h, and T _2F takes a value between 2 and 3 hours;

In the formula: T _steep1 , T _steep2 ,...,T _steepn are respectively: the first, second,..., the starting moment of the nth high-level steep rise segment.

The present invention also provides an ultra-short-term peak load forecasting system, which is improved in that: the system includes:

The collection module is used to collect meteorological forecast information;

A forecasting module, used for ultra-short-term load forecasting based on the established peak load forecasting model and the collected meteorological forecast information;

The peak load prediction model includes: peak load index, meteorological factors affecting peak load and the correlation of each meteorological factor; based on the peak load and meteorological factors affecting peak load.

Further: the prediction module further includes:

Establish sub-modules for establishing peak load forecasting models;

The described establishment submodule further includes:

A first determining unit, configured to determine a peak load index;

The second determination unit is configured to analyze the peak load and the meteorological information based on historical meteorological information to determine the correlation of various meteorological factors affecting the peak load;

A construction unit is configured to construct a peak load forecasting model based on the peak load and the meteorological factors affecting the peak load in combination with an L-M neural network.

Further, the first determining unit further includes:

The analysis and extraction subunit is used to determine the peak load based on pre-acquired data and analyze the peak load to extract peak load indicators.

Further, the second determining unit further includes:

A meteorological factor determination subunit is used to determine relevant meteorological factors according to the historical meteorological information;

The analysis subunit is used to analyze the correlation between peak load and meteorological factors;

The verification subunit is used to verify the correlation of the meteorological factors affecting the peak load based on the robust regression model.

Further, the construction unit further includes:

The first selection subunit is used to select training samples of the L-M neural network based on the peak load and the meteorological factors affecting the peak load;

The third determining subunit is used to determine the data type vector of the comprehensive data sample;

Build subunits for building peak load forecasting models based on training samples.

Further, the first selection subunit is also used to: select system load samples and corresponding comprehensive data samples of k working days or holidays before the forecast date;

Further, the prediction module also includes:

The second selection sub-module is used to select the vector input to the peak load forecasting model;

The calculation sub-module is used to calculate the peak load forecast error until the iteration condition is met;

The output sub-module is used to output peak load forecast results.

Further, the second selection submodule is also used for:

Taking the time threshold as each interval period of the peak load, the prediction object is the set point load of a day; for a single training, the load at the predicted time of m peak samples is used as the load input vector of the neural network;

According to the correlation analysis between the peak load and meteorological factors, the temperature t _k at the same time, the rainfall r _k at the same time, the maximum daily temperature t _max , and the minimum daily temperature t _min of m peak load samples are selected as the peak load forecast The meteorological factors input to the model.

Further, the time threshold is selected as 5 minutes; the set point load is selected as 288 points.

Compared with the closest prior art, the technical solution provided by the present invention has the beneficial effects of:

(1) According to the numerical weather forecast information, the change law of the peak load which is sensitive to the weather is analyzed, and the correlation analysis between the peak load and the meteorological factors is carried out, and the positive correlation between the peak load and the temperature is obtained. For the load data of the power grid system, the high-level steep rise rate of load and the initial moment of high-level steep rise of load are taken as the characteristics of the peak operation mode to reflect and deal with the contradiction that the system response ability does not adapt to the high-level steep rise rate. These two indicators are important for improving peak load Prediction accuracy plays an important role.

(2) On the basis of the establishment of the peak load model, use the L-M neural network method to establish a peak load forecasting method based on meteorological factors, and complete the design of the forecasting model solution algorithm. In view of the lack of data processing ability and convergence difficulty of the original network, this method improves the structure and learning algorithm of the original neural network, replaces the original gradient descent method with the L-M algorithm, improves the dynamic performance of the network, reduces the learning time, and improves the convergence speed. Faster, improved prediction accuracy.

Description of drawings

Fig. 1 is the typical daily load graph of a certain province summer in 2014 provided by the present invention;

Fig. 2 is the peak load curve figure on June 30, 2014 provided by the present invention;

Fig. 3 is the summer peak load and peak time temperature relationship diagram provided by the present invention;

Fig. 4 is the detailed flowchart of ultra-short-term peak load forecasting provided by the present invention;

Fig. 5 is the C-mean fuzzy cluster classification result figure provided by the present invention;

Fig. 6 is the peak load forecast on July 2 provided by the present invention and the actual curve comparison chart;

Fig. 7 is the July 2nd L-M and BP algorithm prediction error comparative figure provided by the present invention;

Fig. 8 is a flow chart of ultra-short-term peak load forecasting provided by the present invention.

Detailed ways

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings.

The following description and drawings illustrate specific embodiments of the invention sufficiently to enable those skilled in the art to practice them. Other embodiments may incorporate structural, logical, electrical, process, and other changes. The examples merely represent possible variations. Individual components and functions are optional unless explicitly required, and the order of operations may vary. Portions and features of some embodiments may be included in or substituted for those of other embodiments. The scope of embodiments of the present invention includes the full scope of the claims, and all available equivalents of the claims. These embodiments of the present invention may be referred to herein, individually or collectively, by the term "invention", which is for convenience only and is not intended to automatically limit the application if in fact more than one invention is disclosed The scope is any individual invention or inventive concept.

Embodiment one,

The present invention provides a method for ultra-short-term peak load forecasting, the flow chart of which is shown in Figure 8, comprising the following steps:

Collect weather forecast information;

Perform ultra-short-term load forecasting based on the established peak load forecasting model and the collected meteorological forecast information;

The peak load prediction model includes: peak load index, meteorological factors affecting peak load and the correlation of each meteorological factor; based on the peak load and meteorological factors affecting peak load.

Among them: the established peak load forecasting model, including:

Step 1: Determine the peak load index;

1.1 Analysis of peak load characteristics

With the growth of economy and the improvement of people's living standards, the power load is increasing year by year. Due to the existence of cooling load in summer, the load value in summer is much higher than that in other seasons. This paper focuses on the weather-sensitive summer peak load characteristics, taking the summer load data of a provincial power grid from June to August 2014 as an example, and excluding holidays and weekend data. Attached Figure 1 is the load curve of a typical day of a provincial power grid in summer. It can be seen from the figure that the province's load is characterized by three peaks and three valleys, which are morning peak (11:15), afternoon peak (17:00), evening peak (21 :15), night trough (4:30), lunch break trough (12:45), evening work trough (19:00).

1.2 Peak load index

The peak load curve on June 30, 2014 is shown in Figure 2. The object of this study is the peak load, and the valley load mode can be treated as a peak (negative valley load) mode. Analyze the load of 24 hours a day, extract the peak load value, peak time, peak-to-valley difference, high load steep rise rate, load high steep rise start time, peak duration and other six characteristics, forming peak load characteristics vector of metrics. The specific meanings of these six features are defined as follows:

(1) Peak and valley load values:

The peak load value is the largest load value P _peak in a day; the valley load value is the smallest load value P _valley in a day.

(2) Peak moment:

The peak hour is the moment when the peak load peak occurs.

(3) Peak-to-valley difference:

The peak-to-valley difference is the difference between the maximum load value and the minimum load value in a day.

P _difference = P _peak – P _valley (1)

(4) Steep rise rate of high load:

The "high level" standard here needs to be clarified. In general, the median value of the daily load curve peak load value and waist load value is bounded. That is, the system load Ph satisfies the relationship:

P _h ≥ P _peak – P _difference /4 (2)

At this point the load level is "high". In special cases, it may not be limited by the above formula.

Define s as the load slope and d as the load time interval, then the load slope at time t is:

The load slope of the high-level steep rise section must reach the steep rise level, and the steep rise level is measured by the following formula:

s _steep ≥0.75s _max (4)

In the formula, s _max is the slope of the steepest period. The ratio of the load increment to the time increment of the high-level steep rise section that satisfies formula (4) is defined as the load high-level steep rise rate.

(5) The starting time of high load steep rise:

The starting moment of the first high-level steep rise period determined according to formula (2) is called the initial high-level steep rise moment of the peak load curve. Suppose a load sample curve is determined by formulas (2), (4), the first one, the second one,..., and the starting time of the nth high-level steep rise section are respectively T _steep1 , T _steep2 ,...,T _steepn , then the starting time of the sample curve’s high steep rise is:

Tsteep＝min(T _steep1 ,T _steep2 ,...,T _steepn ) (5)

(6) Peak duration:

The peak duration is the time span of the load that reaches the high load level; among them: s _t is the load slope at time t; s _steep is the steep rise level; P _difference is the peak-to-valley difference; s _max is the slope of the steepest period; _steep is the starting moment of the steep rise of the sample curve.

Step 2: Analyze and verify the correlation between peak load and meteorological factors;

In recent years, the maximum load of power grid power supply has been increasing year by year, and the peak load is very sensitive to meteorological changes, especially to changes in meteorological factors such as temperature and humidity. Weather changes lead to electricity loads caused by changes in human comfort. The growth of this part of the peak load is very sensitive to climate changes, which is called climate-sensitive peak load. This section analyzes the correlation between peak load and meteorological factors, finds out the key factors affecting meteorological sensitive peak load, and establishes a regression model through robust regression method for verification.

2.1 Correlation analysis

Correlation coefficient is an important index to describe the degree of linear correlation between vectors. The formula for calculating the correlation coefficient is as follows:

In the formula: R is the correlation coefficient; cov(X,Y) is the covariance of X and Y; are the mean square errors of X and Y, respectively.

The instantaneous wind speed, instantaneous temperature, instantaneous air pressure, maximum air pressure, maximum wind direction, maximum wind direction, rainfall, maximum wind speed, instantaneous wind direction, instantaneous humidity, minimum humidity, maximum temperature, minimum temperature, minimum air pressure, maximum Wind speed and other information, as well as the province's three peaks and three valley loads within a month, calculate the correlation coefficients between the three peaks and three valley loads and these meteorological factors. Attached Table 1 lists the meteorological factors that are highly correlated with the load.

It can be seen from Table 1 that the summer peak load in the province is positively correlated with temperature and negatively correlated with rainfall. in:

(1) The three-peak load is sensitive to temperature, and the correlation coefficients are all above 0.80. The correlation coefficient between the morning peak (11:15) load and the time temperature is the largest (0.855), the correlation coefficient between the evening peak (21:15) and the highest temperature is relatively large (0.8069), and the correlation coefficient between the afternoon peak (17:00) and the minimum temperature The coefficient is large (0.8623).

(2) Secondly, the correlation coefficient between load and temperature in midday valley (12:45) is relatively large (0.8236);

(3) The peak-valley load is sensitive to rainfall and is negatively correlated, that is, when the rainfall increases, the load decreases. The correlation coefficients of the afternoon peak (17:00) and afternoon valley (19:00) to rainfall reached -0.537 and -0.7106, respectively.

2.2 Robust regression verification

Consider a robust regression model:

Y=Xβ+ε (7)

Among them, Y=(y _i ) _m×1 is the historical load data at a certain moment in m days; X=(x _ij ) _m×n is the historical temperature data at a certain moment in m days; β=(β _j ) _n×1 is the estimated unknown The parameter vector, ε=(ε _i ) _m×1 is an unobservable random error vector.

Parameter estimation is performed on β=(β _j ) _n×1 in the robust regression model to minimize the following objective function.

Finding the Best Estimated Parameters The key is to determine the weighting function matrix. Let the m-order full-rank diagonal matrix W=(w _i ) _m×m be the weighting function matrix, which can be obtained by solving Solving formula:

in, r _i =y _i -x _(i,j) β _j is the residual term.

The peak load regression model is established with the robust regression method, and the regression equation between the peak load at 11:15 and the temperature at this time is obtained as:

Y＝132.72X+595.238 (10)

As shown in Figure 3, the scatter diagram shows that the relationship between peak load and temperature fits well, with a significant linear correlation, and the peak load increases with the increase of temperature.

Step 3: Construct the peak load forecasting model of the L-M neural network and perform load forecasting:

3.1 LM neural network

The L-M (Levenberg-Marquardt) neural network method combines the advantages of the gradient descent method and the Gauss-Newton method. The advantage is that when the number of network weights is small, the convergence is very fast, the learning time is significantly shortened, and it is obviously superior in terms of training times and accuracy. than the ordinary BP algorithm.

In the process of training the network, the error between the network output and the ideal is:

E(t _i )=Y(t _i )-A(t _i ), i=1,...,m (11)

By adjusting the value in the weight matrix W, the error between the output of the network and the ideal value is minimized, that is, the following formula is satisfied:

Let W represent a vector composed of weights and thresholds, and the dimension of the vector is l. The L-M algorithm is an improved Gauss-Newton method, and its form is:

ΔW＝(J ^T J+μI) ^-1 J ^T E (13)

in:

Set the prediction accuracy threshold ε, and if formula (15) is satisfied, the iteration will end. Otherwise, modify the weights of neurons in each layer until it is satisfied.

Among them: E(t _i ) is the error between the network output and the ideal, Y(t _i ) is the network output, that is, the power, A(t _i ) is the ideal value, W ^opt is the minimum error between the network output and the ideal value The optimal W value, ΔW is a vector composed of weights and thresholds obtained by the improved Gauss-Newton method, W is the optimal value that minimizes the error between the output value of the peak load forecasting model and the ideal value, and J is the Jacobian matrix, J ^T is the transpose of the Jacobian matrix, E is E(W ^opt ) indicating the minimum error between the output of the network and the ideal value, and m is the number of peak load samples in the model.

3.2 Peak load forecasting model of L-M neural network:

3.2.1 Selection of samples:

This paper selects system load samples 8-14 working days (or holidays) before the forecast date. In order to ensure that enough samples can be selected, the daily maximum temperature, daily minimum temperature, weather conditions, rainfall, and daily types are selected as comprehensive Data samples to build load patterns.

Let the clustering sample weather comprehensive data type vector be:

D _k ＝(D _k1 D _k2 D _k3 D _k4 D _k5 ) ^T (16)

where k=1,2,...,p. p is the number of days selected, D _k1 is the daily maximum temperature on the k-th day, D _k2 is the daily minimum temperature on the k-th day, D _k3 is the weather condition on the k-th day, D _k4 is the average humidity on the k-th day, D _k5 The day type for the kth day.

After the C-means fuzzy clustering method, the iterative optimization calculation is used to obtain the vectors belonging to the same class. In this class, the peak load sample curves of the days corresponding to the vectors are selected, which are respectively recorded as Y _p1 (t), Y _p2 (t) ,…,Y _pm (t), and thus form a peak load pattern, namely:

M _p = {Y _p1 (t), Y _p2 (t),..., Y _pm (t)} (17)

Where m is the number of peak load samples in the model.

3.2.2 Selection of input vector

Take 5 minutes as each interval of peak load, so the forecast object is 288 points of load in one day. For a single training session, the predicted time load of m peak samples is used as the load input vector of the neural network.

According to the correlation analysis between peak load and meteorological factors in Section 2.1, the simultaneous temperature t _k , simultaneous rainfall r _k , sample daily maximum temperature t _max , and sample daily minimum temperature t _min of m peak load samples are selected as Meteorological factor input to neural network. where k=1,2,...,m.

3.3 Ultra-short-term peak load forecasting process

The flow of the ultra-short-term peak load forecasting algorithm is shown in Figure 4, and the steps are as follows:

Step 1: Select the forecast date;

Step 2: According to meteorological factors, select the comprehensive data type D _k , and perform the maximum and minimum normalized data types to obtain D′ _k ;

Step 3: Using C-means fuzzy clustering, select m peak load curves M _p = {Y _p1 (t), Y _p2 (t), ..., Y _pm (t)} that are close to the forecast day as the peak load curves belonging to the same peak load a sample of the pattern;

Step 4: Input the neural network vector according to the selection method of the input vector.

Step 5: set up a multi-layer feed-forward L-M network prediction model, and calculate the prediction error until formula (15) is satisfied;

Step 6: Output load forecast power estimates

Step 7: For n peak load samples whose forecast category matches the forecast day and belong to the same model, calculate the daily high and steep start time vector T _steep(1×n) = (T _steep1 T _steep2 …T _{steep n} ) ^T , and then calculate their average starting time T _Avsteep ;

Step 8: Determine the peak time period [T _AVsteep -T _2H , T _AVsteep +T _2F ], where T _2H can take a value from 0.5 to 1 hour, and T _2F can take a value from 2 to 3 hours.

Embodiment two,

Taking the comprehensive data of meteorological factors and sample data of peak load during the 2014 summer solstice of a provincial grid as an example, the peak load of the power grid on July 2, 2014 is predicted by using the LM neural network peak prediction algorithm. Firstly, 11 samples of weekdays closer to July 2 are selected for clustering. The comprehensive data type samples are shown in Table 2. In Table 2, type 1 represents a weekday, and the difference coefficient of weather conditions is shown in Table 3. Using the C-mean fuzzy clustering method, the results are shown in Figure 5. In this two-category result graph, select the load sample curves of the day corresponding to the same category ⊕ symbol, and form a load pattern M _p {X _p (t),X _p (t),…,X _p (t) }, where the components respectively represent the peak load sample curves on June 17, June 21, June 26, June 27, June 28 and July 1.

According to the ultra-short-term peak load forecasting process, the ultra-short-term peak load on July 2, 2014 is predicted, and the L-M forecast result is obtained. The L-M predicted and actual value curves are shown in Figure 6.

Compare the L-M prediction results with the BP neural network prediction results, and calculate the average relative error. The error comparison is shown in Figure 7. It can be seen from the figure that the L-M algorithm predicts that the peak duration interval is [19:30, 22:25]; the average relative error is 0.61785%; the maximum relative error is 1.7083%; the predicted peak load peak value is 7667MW, and the actual peak load peak value is 7694.7MW, the relative error of the peak value is 0.3599%; the start time of the predicted high steep rise is 19:35. The number of network learning training is 14 times. The common BP method predicts that the average relative error is 1.64205%, the maximum relative error is 3.4409%, the peak relative error is 1.765%, and the number of network learning and training is 1000 times.

Table 1 Correlation coefficient between three peaks and three valley loads and meteorological factors in summer 2014

Table 2 Comprehensive data type sample

Table 3 Difference coefficient table of weather conditions

The ultra-short-term peak load forecasting method based on L-M neural network and numerical weather forecast proposed by the present invention establishes a peak load pattern with characteristic vectors such as peak start time and high steep rise rate, and the abundant meteorological information collected through numerical weather forecast, The correlation between peak load and meteorological factors was analyzed and verified by robust regression method, and the key meteorological factors affecting peak load were found. Then the ultra-short-term load forecasting modeling is carried out by L-M neural network method. The peak load samples are selected based on the C-means fuzzy clustering method, and the key meteorological factors and load vectors of the peak load samples are selected as the input of the neural network for training and prediction. The prediction method and technical scheme are the core content of this technology.

Figure 3 shows the relationship between summer peak load and peak temperature, Figure 4 shows the flow chart of ultra-short-term peak load forecasting, Figure 5 shows the results of C-mean fuzzy clustering classification, Figure 6 shows the comparison between peak load forecast and actual curve Figure 7 and the L-M and BP algorithm prediction error comparison chart; and the correlation coefficient table between three peaks and three valley loads in summer and meteorological factors as shown in Table 1, the comprehensive data type sample table in Appendix 2, and the weather condition difference coefficient table in Appendix 3.

Embodiment three

The present invention also provides an ultra-short-term peak load forecasting system, which is improved in that: the system includes:

The collection module is used to collect meteorological forecast information;

A forecasting module, used for ultra-short-term load forecasting based on the established peak load forecasting model and the collected meteorological forecast information;

The peak load prediction model includes: peak load index, meteorological factors affecting peak load and the correlation of each meteorological factor; based on the peak load and meteorological factors affecting peak load.

The prediction module further includes:

Establish sub-modules for establishing peak load forecasting models;

The described establishment submodule further includes:

A first determining unit, configured to determine a peak load index;

The second determination unit is configured to analyze the peak load and the meteorological information based on historical meteorological information to determine the correlation of various meteorological factors affecting the peak load;

A construction unit is configured to construct a peak load forecasting model based on the peak load and the meteorological factors affecting the peak load in combination with an L-M neural network.

The first determining unit further includes:

The first determining subunit is configured to determine the peak load based on pre-acquired data and analyze the peak load to extract the peak load index.

The second determining unit further includes:

The second determination subunit is used to determine relevant meteorological factors according to the historical meteorological information;

The analysis subunit is used to analyze the correlation between peak load and meteorological factors;

The verification subunit is used to verify the correlation of the meteorological factors affecting the peak load based on the robust regression model.

Described construction unit further comprises:

The first selection subunit is used to select training samples of the L-M neural network based on the peak load and the meteorological factors affecting the peak load;

The third determining subunit is used to determine the data type vector of the comprehensive data sample;

Build subunits for building peak load forecasting models based on training samples.

The first selection subunit is further used to: select system load samples and corresponding comprehensive data samples of k working days or holidays before the forecast date;

The prediction module also includes:

The second selection unit is used to select the vector input to the peak load forecasting model;

A calculation unit, used to calculate the peak load forecast error until the iteration condition is satisfied;

The output unit is used to output peak load forecast results.

Further, the second selection unit is also used for:

Taking the time threshold as each interval period of the peak load, the prediction object is the set point load of a day; for a single training, the load at the predicted time of m peak samples is used as the load input vector of the neural network;

According to the correlation analysis between the peak load and meteorological factors, the temperature t _k at the same time, the rainfall r _k at the same time, the maximum daily temperature t _max , and the minimum daily temperature t _min of m peak load samples are selected as the peak load forecast The meteorological factors input to the model. The time threshold is selected as 5 minutes; the set point load is selected as 288 points.

The present invention proposes an ultra-short-term peak load forecasting method and its system. The peak load pattern is established with the peak start time and the peak steep rise rate and other characteristic vectors. The rich meteorological information collected by the numerical weather forecast can predict the peak load and weather The correlation analysis of the factors was carried out and verified by the robust regression method, and the key meteorological factors affecting the peak load were found. Then the ultra-short-term load forecasting modeling is carried out by L-M neural network method. The peak load samples are selected based on the C-means fuzzy clustering method, and the key meteorological factors and load vectors of the peak load samples are selected as the input of the neural network for training and prediction. Actual examples prove that this technology is effective for peak forecasting and improves the forecasting accuracy of ultra-short-term load.

The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art can still modify or equivalently replace the specific embodiments of the present invention. , any modifications or equivalent replacements that do not deviate from the spirit and scope of the present invention are within the protection scope of the claims of the present invention pending application.

Claims (23)

1. A method for ultra-short term peak load prediction, characterized by:

collecting weather prediction information;

ultra-short-term load prediction is carried out based on a pre-established peak load prediction model and the collected meteorological prediction information;

the peak load prediction model comprises: peak load index, meteorological factors affecting peak load and correlation of each meteorological factor; based on the peak load and meteorological factors affecting the peak load.

2. The ultra-short term peak load prediction method of claim 1, wherein the pre-established peak load prediction model comprises:

determining a peak load index;

analyzing the peak load and the weather information based on historical weather information to determine the correlation of weather factors influencing the peak load;

and constructing a peak load prediction model by combining an L-M neural network based on the peak load and the meteorological factors influencing the peak load.

3. The ultra-short term peak load prediction method of claim 2, wherein said determining a peak load indicator comprises:

determining peak load based on pre-acquired data and analyzing the peak load to extract a peak load index;

the peak load indicators include:

1) peak top and valley bottom load values:

peak top load value P_peakIs the maximum load value during the day; load value P of valley bottom_valleyIs the minimum load value during the day;

2) the peak top time:

the peak-top time is the time when the peak load peak occurs;

3) peak-to-valley difference:

the expression for the peak-to-valley difference is as follows:

P_difference＝P_peak–P_valley(1)

wherein, P_differenceIs the peak-to-valley difference;

4) a high end load ramp rate comprising:

system load P_hSatisfies the relationship:

P_h≥P_peak–P_difference/4 (2)

the load level is at a high level at this time;

the high-end load slope at time t is:

in the formula: s is the high load slope, d is the high load time interval, P_tIndicating the high level load value, P, at time t_tRepresenting the high-order load value at the time t-1; s_tIs the high load slope at time t;

the load slope of the high steep rise section must reach a steep rise level, measured as:

s_steep≥0.75s_max(4)

in the formula: s_maxThe slope of the steepest time period;

5) load high-position steep rise starting time:

the expression is as follows:

T_steep＝min(T_steep1,T_steep2,…,T_steepn) (5)

wherein s is_steepIs at a steep level; t is_steep1,T_steep2,…,T_steepnRespectively as follows: no. 1, No. 2, No. n high steep rising paragraph start time; t is_steepDefining the ratio of the load increment and the time increment of the high-position steep rising section satisfying the formula (4) as the high-position steep rising speed of the load for the high-position steep rising starting moment of the sample curve;

6) peak duration:

peak duration is the time span made up of loads reaching high load levels.

4. The ultra-short term peak load forecasting method of claim 2, wherein said analyzing the peak load and weather information based on historical weather information to determine the relevance of weather factors affecting peak load comprises:

determining related meteorological factors according to the historical meteorological information;

analyzing the correlation of the peak load and meteorological factors;

and verifying the relevance of each meteorological factor influencing the peak load based on a robust regression model.

5. The ultra-short term peak load prediction method of claim 4, wherein the meteorological factors comprise: instantaneous wind speed, instantaneous temperature, instantaneous air pressure, maximum wind direction, rainfall, maximum wind speed, instantaneous wind direction, instantaneous humidity, minimum humidity, maximum temperature, minimum air pressure, maximum wind speed information.

6. The ultra-short term peak load prediction method of claim 5, wherein said analyzing the correlation of peak load to meteorological factors comprises: calculating the correlation coefficient of each meteorological factor based on the following formula:

in the formula: r is a correlation coefficient; cov (X, Y) is the covariance of X and Y;mean square error of X and Y, respectively; x is historical load data; y is: historical temperature data.

7. The ultra-short term peak load prediction method of claim 4, wherein the robust regression model comprises: constructing a robust regression model calculation formula as follows:

Y＝Xβ+ε (7)

wherein: y ═ Y_i)_m×1Historical load data at a certain moment of m days; x ═ X_ij)_m×nHistorical temperature data at a time of m days, β ═ β_j)_n×1Is an estimated unknown parameter vector, e ═ e_i)_m×1Is a random error vector that is not observable;

for β in the robust regression model (β) based on the following formula_j)_n×1And (3) performing parameter estimation:

wherein D (β) represents the objective function in the robust regression model, y_iRepresents the historical load data at a certain time of the ith day, i is 1, …, m, x_ijThe temperature data of a certain time on the ith day is shown, and m represents the total number of days;

finding optimal estimation parametersThe key point of (1) is to determine a weighting function matrix; let m-order full-rank diagonal matrix W be (W)_i)_m×mFor weighting function matrix, obtaining optimal estimation parameters by solvingSolving the formula:

wherein,r_i＝y_i-x_(i,j)β_jis a residual term; x^TIs transposed data of the historical temperature data X at a certain moment of m days.

8. The ultra short term peak load prediction method of claim 2, wherein said building a peak load prediction model in conjunction with an L-M neural network based on said peak load and said meteorological factors affecting peak load comprises:

selecting a training sample of the L-M neural network based on the peak load and the meteorological factors influencing the peak load;

determining a data type vector of the synthetic data sample;

and constructing a peak load prediction model according to the training samples.

9. The method of ultra-short term peak load prediction according to claim 8, wherein said selecting training samples of L-M neural networks based on said peak load and said meteorological factors affecting peak load comprises: selecting system load samples and corresponding comprehensive data samples of k working days or holidays before a prediction day;

the synthetic data type vector of the synthetic data sample is:

D_k＝(D_k1D_k2D_k3D_k4D_k5)^T(16)

wherein: k denotes the number of days k 1,2, …, p, p being the number of selected days, D_k1Day maximum temperature on day k, D_k2Day minimum temperature on day k, D_k3Weather conditions on day k, D_k4Average humidity of day k, D_k5A day type for day k;

constructing a peak load prediction model according to the training samples, comprising:

performing iterative optimization calculation to obtain vectors belonging to the same class from the comprehensive data type vectors of the comprehensive data samples by a C-means fuzzy clustering method, and selecting the peak load sample curves of the days corresponding to the vectors in the class and respectively recording the curves as Y_p1(t), Y_p2(t),…,Y_pm(t)；

Determining a peak load prediction model according to the peak load sample curve, wherein the expression is as follows:

M_p＝{Y_p1(t),Y_p2(t),…,Y_pm(t)} (17)

wherein: y is_p1(t),Y_p2(t),…,Y_pm(t) a first, second, and/or mth peak load sample curve, respectively; m is the number of peak load sample curves.

10. The ultra-short term peak load prediction method of claim 1, wherein ultra-short term load prediction based on the peak load prediction model in combination with near weather prediction comprises:

selecting a vector input to a peak load prediction model;

calculating a peak load prediction error until an iteration condition is met;

and outputting a peak load prediction result.

11. The ultra short term peak load prediction method of claim 10, wherein said selecting a vector input to a peak load prediction model comprises:

taking a time threshold as each interval period of peak load, then predicting that the object is a set point load for a day; for a single training, taking the predicted moment loads of m peak samples as load input vectors of a neural network;

according to the correlation analysis between the peak load and the meteorological factors, the simultaneous temperature t of m peak load samples is selected_kAnd simultaneously the rainfall r_kMaximum daily temperature t of the sample_maxDay minimum temperature t of the sample_minAs meteorological factor input for a peak load prediction model.

12. The ultra short term peak load prediction method of claim 11, wherein said time threshold is selected to be 5 minutes; the setpoint load takes 288 points.

13. The ultra short term peak load prediction method of claim 10, wherein the iteration condition is represented by the following equation:

wherein: y (t)_i) Predicting the model output value, A (t), for peak loads_i) The peak load prediction model is an ideal value, W is an optimal value which enables the error between the output value of the peak load prediction model and the ideal value to be minimum, and epsilon is a prediction precision threshold value; i denotes day i.

14. The ultra-short term peak load prediction method of claim 10, wherein the peak load prediction result comprises:

load forecast power estimation value

Calculating high-order steep rising initial time vectors T of various days for n peak load samples with the prediction types consistent with the prediction days and belonging to the same mode_steep(1×n)＝(T_steep1T_steep2…T_steepn)^TCalculating the average starting time T_Avsteep；

Peak time [ T_AVsteep-T_2H，T_AVsteep+T_2F]Wherein T is_2HThe value is between 0.5 and 1h, T_2FTaking values within 2-3 h;

in the formula: t is_steep1,T_steep2,…,T_steepnRespectively as follows: no. 1, No. 2, No. n high steep rising paragraph start time.

15. An ultra-short term peak load prediction system, comprising: the system comprises:

the acquisition module is used for acquiring meteorological prediction information;

the prediction module is used for carrying out ultra-short-term load prediction based on the built peak load prediction model and the collected meteorological prediction information;

the peak load prediction model comprises: peak load index, meteorological factors affecting peak load and correlation of each meteorological factor; based on the peak load and meteorological factors affecting the peak load.

16. The ultra short term peak load prediction system of claim 15, wherein: the prediction module further comprises:

the building submodule is used for building a peak load prediction model;

the establishing submodule further includes:

a first determining unit for determining a peak load index;

the second determining unit is used for analyzing the peak load and the weather information to determine the correlation of each weather factor influencing the peak load based on historical weather information;

and the construction unit is used for constructing a peak load prediction model by combining an L-M neural network based on the peak load and the meteorological factors influencing the peak load.

17. The ultra-short term peak load prediction system of claim 16, wherein the first determination unit further comprises:

and the analysis and extraction subunit is used for determining the peak load based on the pre-acquired data and analyzing the peak load to extract the peak load index.

18. The ultra-short term peak load prediction system of claim 16, wherein the second determination unit further comprises:

the meteorological factor determining subunit is used for determining related meteorological factors according to the historical meteorological information;

the analysis subunit is used for analyzing the correlation between the peak load and the meteorological factor;

and the verification subunit is used for verifying the correlation of each meteorological factor influencing the peak load based on the robust regression model.

19. The ultra-short term peak load prediction system of claim 16, wherein the building unit further comprises:

the first selection subunit is used for selecting a training sample of the L-M neural network based on the peak load and the meteorological factors influencing the peak load;

a third determining subunit, configured to determine a data type vector of the synthetic data sample;

and the construction subunit is used for constructing a peak load prediction model according to the training samples.

20. The ultra-short term peak load prediction system of claim 19, wherein said first selected subunit is further configured to: and selecting system load samples and corresponding comprehensive data samples of k working days or holidays before the prediction day.

21. The ultra-short term peak load prediction system of claim 15, wherein the prediction module further comprises:

a second selection submodule for selecting a vector input to the peak load prediction model;

the calculation submodule is used for calculating a peak load prediction error until an iteration condition is met;

and the output submodule is used for outputting a peak load prediction result.

22. The ultra-short term peak load prediction system of claim 21, wherein said second selection submodule is further configured to:

taking a time threshold as each interval period of peak load, then predicting that the object is a set point load for a day; for a single training, taking the predicted moment loads of m peak samples as load input vectors of a neural network;

according to the correlation analysis between the peak load and the meteorological factors, the simultaneous temperature t of m peak load samples is selected_kAnd simultaneously the rainfall r_kMaximum daily temperature t of the sample_maxDay minimum temperature t of the sample_minAs meteorological factor input for a peak load prediction model.

23. The ultra short term peak load prediction system of claim 22, wherein said time threshold is selected to be 5 minutes; the setpoint load takes 288 points.