Rainfall forecasting using artificial neural networks

Rainfall is a complex meteorological process that affects the environment, human based activities, agriculture, transportation, and almost every aspect of life. The ability to determine the amount of rain that will fall is helpful for different agricultural industries. It enhances decision-making and management of farming processes, including plantation, irrigation, fertilization, and harvest. Different rainfall forecasting methods have been suggested and used to forecast rainfall over various durations. The existing forecasting models release forecasts over large grid areas, and the forecasted output is given to end-users as a probabilistic value which is considered uninformative because it does not provide accurate information about the type of expected rainfall. These models have also shown low prediction accuracy on some occasions. In this thesis, new approaches are proposed for predicting monthly rainfall. The approaches are based on single artificial neural networks and ensembles of artificial neural networks. The first approach proposes an evolutionary algorithm to select the most significant features for single neural networks. The second approach extends the first approach by including network parameters in the selection process. The third approach introduces a fusion of multiple single neural networks and develops novel ensembles of neural networks for rainfall prediction. Several fusion methods are proposed to combine the single neural networks. The fourth approach uses ensemble components selection while building the rainfall prediction model. Two types of forecasts were targeted in this thesis: numerical and categorical. Numerical prediction is the process of predicting the actual amount of rainfall. In categorical prediction, rainfall is classified into categories, such as above average and below average. The proposed approaches were applied in order to predict rainfall for areas of eastern Australia. The datasets were created by collecting and processing weather variables from multiple online resources. Several local and global weather attributes were used as possible predictors of rain. These predictors included temperature, solar attribute, and climate indices. A climate index represents a particular phenomenon over a selected area in oceans including Pacific and Indian oceans. Australian rainfall variability is correlated to these local and global weather attributes. Various statistical measurements, including Mean Absolute Error, Root Mean Square Error, Pearson Correlation, Skill Scores, classification error, and f-score, were used to assess the performance of the proposed approaches. The proposed approaches were compared to alternative techniques, and better performance was recorded with proposed neural network based approaches. In addition, the developed models using the proposed approaches were compared with the new forecasting system released by the Bureau of Meteorology, and better performance was obtained with the developed models used in this thesis. The results and comparative analysis in this thesis show that the proposed neural network based ensemble approach is appropriate for monthly rainfall prediction. It was found that input features and neural network parameters should be carefully chosen when designing a neural network. In addition, correct selection of ensemble components increases the efficiency of the ensemble model. The proposed models can be useful for agricultural industries such as sugarcane, wheat, and cotton. Because various industries record rainfall and weather data, the proposed approaches can be utilized to build forecasting models for these industries, as adequate records of related atmospheric data are available on multiple online resources.