Habilitation thesis - abstract

In the habilitation thesis the models for short-term electric load forecasting based on the methods of machine learning, pattern recognition and computational intelligence are proposed. The common features of these models are learning on examples and using the similarities between patterns of seasonal cycles of the load time series. These series are non-stationary, heteroscedastic, show trend, many cycles of seasonal fluctuations and random noise. The new approach based on the pattern similarity simplify the forecasting problem and enables the development of effective forecasting models. These models are based on the following assumption: if the patterns of the seasonal cycles of the time series are similar to each other (input patterns), then the patterns of cycles following them (forecast patterns) are also similar to each other. This assumption allows us to build forecasting models using analogies between repetitive fragments of the time series with seasonal fluctuations. In this work a method of statistical analysis of the similarities between patterns, which allows us to confirm the validity of the assumption is proposed. Many definitions of input and output patterns and measures of similarity between them are given.
The forecasting machine learning models based on similarities between pattern of the daily cycles of time series are proposed:

  • model based on the kernel estimators,
  • models based on the nearest neighbor estimators: k nearest neighbor estimator and fuzzy estimators of the regression function,
  • model based on artificial immune system with local feature selection,
  • models based on pattern clustering: hard and fuzzy k-means, self-organizing feature map, neural gas, hierarchical clustering and two new artificial immune systems.

For learning and optimization of the models deterministic and stochastic methods, such as evolutionary algorithms and original tournament searching method, are used. The following analyses of the models are performed: computational complexity, sensitivity to changes in parameter values and resistance to noisy and incomplete data. The methods of statistical analysis of forecasting errors are described. The accuracies of the proposed models and classical models (ARIMA, exponential smoothing), as well as the neural network (multilayer perceptron) are compared.
Models based on the pattern similarity are distinguished by simplicity, clear structure and simple principle of operation. The number of parameters here is small, which implies a less complex process of learning and optimization, and better generalization of the model. The proposed models use nonparametric local regression, which ensures high accuracy, competitive with the accuracy of other forecasting models dedicated to this problem.