Loading...

Bukan Reservoir Station Inflow and Rainfall Forecasting Model Using Deep Learning Algorithms

Mirzaei, Amir Mohammad | 2019

589 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 52910 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Tajrishi, Masoud
  7. Abstract:
  8. The amount of water entering the dam is one of the types of information that has a direct impact on how to plan the discharge and storage of the dam's water supply. Given the precise input streamflow predictions of the next few days, the corresponding authorities can better manage the amount of water that should remain in the storage in the days to come. In this regard, it is no longer necessary to release huge amounts of stored water due to the lack of information on the volume of water entering. Consequently, the volume of water stored in the reservoir, increases. Also, the probability of occurrence of potential hazards such as dam overflow is prevented.Over the last few years, due to the huge developments of computer science and artificial intelligence, it has been possible to provide models for predicting hydrological parameters, which, although pose more complexity than older ones, have greatly improved the accuracy of computations.Due to the existence of complex and nonlinear relations between the amount of inputs streamflow and the parameters involved, using nonlinear methods will provide outputs closer to reality.In the present study, models with the ability to predict the amount of input streamflow to Boukan dam for the next four days ahead were developed. Models capable of predicting the amount of rainfall for the next day ahead at Boukan dam were also created.Various daily time series data with a length of nearly twenty years were used for the study, which includes river flow and rainfall data at different sites, upstream of the dam and also meteorological information of a synoptic weather reporting station. These data were pre-processed.After the data were prepared, modeling began. Required Python codes were written. For the purpose of finding the best hyperparameters for the models, a grid search method was utilized and several deep learning models such as Long-Short-Term Memory (LSTM) and MultiLayer Perceptron (MLP) models with different architectures were created.Two timing scenarios were defined. The first timing scenario contained the spring seasons of the time series data and for the second scenario, in addition to the data extracted for the first scenario, the last months of the winter seasons were also added to the dataset. For each timing scenario, two methods for choosing the best model architecture among all architectures were applied. These methods required a huge number of models to be created. For each of the scenarios, an approximate number of 600 and 60 models were created to choose the best input streamflow and rainfall predictive model architectures, respectively.The results showed that although the length of the time series data was small, the models performed very well. In particular, in the first scenario, R2 for the predicted input streamflow on the test data, for the first, second, third and fourth days ahead were 0.93, 0.9, 0.86 and 0.85, respectively. In the second scenario, R2 for the predicted input streamflow on the test data, for those days were 0.91, 0.9, 0.87 and 0.85, respectively. In the first and second scenarios, R2 for the predicted rainfall near Boukan dam on the test data, for the first day ahead, was 0.53 and 0.37, respectively.Furthermore, a number of statistical indices like the Nash-Sutcliffe Coefficient of Efficiency (NSCE), Peak Weighted Root Mean Square Error (PWRMSE), Mean absolute error (MAE) and RMSE were calculated. The indices showed promising values indicating the goodness of models created
  9. Keywords:
  10. Machine Learning ; Deep Learning ; Nonlinear Regression ; Stream Flow Forecasting ; Neural Network ; Bookan Dam ; Long Short Term Memory (LSTM) ; Long-Lead Streamflow Forecasting ; Seasonal Rainfall Prediction

 Digital Object List

 Bookmark

No TOC