Reservoir Inflow Forecasting Using Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System Techniques
Googhari, Shahram Karimi (2007) Reservoir Inflow Forecasting Using Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System Techniques. PhD thesis, Universiti Putra Malaysia.
A feed forward Artificial Neural Network (ANN) and an Adaptive Neuro-Fuzzy Inferences System (ANFIS) reservoir inflow models were developed to investigate their potential in forecasting reservoir inflows. The site for the study is the Sembrong dam catchment which is located about 10km from Air Hitam town on the Air Hitam-Kluang road in the state of Johor, with an area of 130 square kilometers. The models consists of 9 inputs (previous last five-day reservoir inflow and last four-day average rainfall across the catchment) and are able to forecast the next day inflow into the reservoir. Average rainfall across the catchment was calculated by Theissen polygons. The 6 years daily data from 1995-1997 and 2002-2004 were used for training and validation of the models. Cross validation of training and validation data sets was also considered to obtain the best data set. Daily reservoir inflow was computed using a water balance equation. The reservoir inflow and rainfall data sets were examined for normal distribution and the best data transformation was used. Autocorrelation, partial autocorrelation and cross correlation functions were used to find the best model inputs. The ANN models were trained and simulated using a written program in MATLAB environment (M-file) with raw and transformed data. The ANFIS models were built using the Fuzzy Toolbox of MATLAB. The Subtractive Clustering (SC) technique was employed to find the optimal number of rules. Different ANFIS structures were constructed by changing the SC parameters. All models were trained by the ANFIS editor of MATLAB with hybrid method. An M-file was written for calculating the different performance criteria of ANFIS models after simulating models during training, validation and testing. After selecting the best ANFIS structure, the response of the model to different types of membership functions was investigated. The models were tested with the 10 months daily data of 2005. The best architecture of the ANN model was a 9-13-1 model which means a model with 9 inputs, 1 hidden layer with 13 neurons and 1 output. The model was trained based on the Leven-berg Marquardt algorithm with sigmoid activation functions. Simulation results for the independent testing data series showed that the model can perform well in simulating peak flows as well as base flows. The ANN model has been constructed for a strong non-linear input/output data. Comparisons of different ANN models for different data sets revealed that cross validation of data was effective in improving models performances. Data pre-processing to transform data to normal distribution before the training, results in better generalization and persistency of ANN models during testing. The ANFIS models were built using the best data subset resulting from ANN modeling. The models were trained with normalized and non-normalized data. The selected ANFIS model was trained with normalized data with 6 Gaussian membership functions for each of 9 inputs and 6 rules. Comparisons of different performances of ANFIS models showed that data normalization can improve the model performances during training and testing. Simulation results for the independent test data series by the ANFIS model showed the ability of this model to forecast daily reservoir inflow in a tropical ungauged catchment. Sensitivity of the ANFIS model using different types of membership functions indicated that the best one is the Gaussian membership function. The simulation results from the selected ANFIS and ANN models during training, validation and testing revealed the superiority of the ANN model. The selected ANFIS model gives lower values in most of the performance indices during training. For validation and testing, all performance indices of selected ANFIS model were inferior to those of the ANN model. The weakness of ANFIS model is shown in its inability to forecast individual peak flows. The sudden flow changes in these small tropical catchments resulting in these peak flows are common due to their small areal extent and to the intense localized phenomenon of tropical showers.
Repository Staff Only: Edit item detail