Forecasting method of load forecasting

Power load forecasting is divided into classical forecasting methods and modern forecasting methods. Trend extrapolation

It predicts the future load situation according to the changing trend of load. Although the power load is random and uncertain, under certain conditions, there is still an obvious changing trend, such as agricultural electricity consumption. In winter, the climatic conditions have not changed much, and the daily electricity consumption is relatively stable, showing a relatively stable trend. This trend can be linear, nonlinear, periodic or aperiodic, and so on.

Time series method

Time series method is one of the most commonly used short-term load forecasting methods. It aims at the characteristics of the random process presented by the whole observation sequence, establishes and estimates the models of the random process that produces the actual sequence, and then uses these models to forecast. It uses the inertia characteristics and time continuity of power load change, and through the analysis and processing of historical data time series, determines its basic characteristics and changing rules, and predicts the future load.

Time series prediction methods can be divided into two categories: deterministic and stochastic. Deterministic time series is used as model residual to estimate the size of prediction interval. The stochastic time series prediction model can be regarded as a linear filter. According to the characteristics of linear filters, time series can be divided into autoregressive (ar), moving average (MA), autoregressive-moving average (ARMA), cumulative autoregressive-moving average (ARIMA) and transfer function (TF), and its load forecasting process is generally divided into five stages: model identification, model parameter estimation, model verification, load forecasting, accuracy test and forecast value correction.

Regression analysis method

Regression analysis is to establish an analyzable mathematical model and predict the future load according to the historical data of the past load. By using the regression analysis method in mathematical statistics, the relationship between variables is determined by analyzing the observed data of variables, so as to realize the prediction. In the late 1980s, some modern forecasting methods based on the theories of emerging disciplines were gradually applied successfully. Among them, there are mainly grey mathematics theory, expert system method, neural network theory and fuzzy prediction theory.

Grey mathematical theory

Grey mathematics theory regards load sequence as a real system output, which is the comprehensive result of many influencing factors. The unknowns and uncertainties of these factors become the grey characteristics of the system. Grey system theory transforms load sequence into regular power generation sequence, and re-models it for load forecasting.

Expert system method

Expert system method is to analyze the load data and weather data in the past few years stored in the database in detail, collect the knowledge of experienced load forecasters and extract relevant laws. With the help of expert system, load forecasters can identify the types of forecasting days, consider the influence of weather factors on load forecasting, and make load forecasting according to certain reasoning.

Neural network theory

Neural network theory is to use the learning function of neural network to let computers learn the mapping relationship contained in historical load data, and then use this mapping relationship to predict the future load. Because this method has strong robustness, memory ability, nonlinear mapping ability and strong self-learning ability, it has a great application market, but its disadvantage is that the learning convergence rate is slow and it may converge to local minimum. And it is difficult to express knowledge and make full use of the fuzzy knowledge existing in the dispatcher's experience.

Fuzzy load forecasting

Fuzzy load forecasting is a research hotspot in recent years.

Fuzzy control is to apply fuzzy mathematics theory to the adopted control method, so that it can carry out deterministic work and effectively control some controlled processes that cannot build mathematical models. No matter how the fuzzy system is calculated, it is a nonlinear function from the point of view of input and output. For any nonlinear continuous function, the fuzzy system is to find a membership function, a reasoning rule and a defuzzification method, so that the designed fuzzy system can arbitrarily approximate this nonlinear function. (1) table search method:

Tabular method is a relatively simple and clear algorithm. The basic idea of this method is to generate fuzzy rules from known input and output data pairs and form a fuzzy rule base, and the final fuzzy logic system will be generated by the combined fuzzy rule base.

This is an easy-to-understand algorithm, because it is a sequential generation process and does not need repeated learning. Therefore, this method also has a great advantage that the fuzzy system is superior to the neural network system, that is, the construction is simple and fast.

(2) Takagi-Kano fuzzy prediction algorithm based on neural network integration;

It uses neural network to obtain the joint membership function of conditional partial input variables. The function f(X) of the conclusion part can also be expressed by neural network. The neural network adopts forward BP network.

(3) Improved fuzzy neural network model algorithm:

Fuzzy neural network is a global approximator. There seems to be a natural connection between fuzzy systems and neural networks. Fuzzy neural network is essentially the realization of fuzzy system, which is to give fuzzy input signals and fuzzy weights to conventional neural networks (such as feedforward neural networks and HoPfield neural networks).

There are many methods for modeling complex systems, and good application results have been achieved. However, the main disadvantages are low model accuracy and long training time. The model of this method has obvious physical significance, high accuracy and fast convergence, and it belongs to an improved algorithm.

(4) Back propagation learning algorithm:

The application of fuzzy logic system mainly lies in that it can be used as a model of nonlinear systems, including those with human operators. Therefore, it is very important to study the nonlinear mapping ability of fuzzy logic system in the sense of function approximation. Function approximation means that a fuzzy logic system can uniformly approximate any nonlinear function defined on the dense set with any precision, and its advantage lies in the systematic and effective use of language information. The universal approximation theorem shows that there must be such a Gaussian fuzzy logic system, which can approximate any given function with arbitrary precision. Back propagation BP learning algorithm is used to determine the parameters of Gaussian fuzzy logic system, and the identified model can approach the real system well, thus improving the prediction accuracy.