Application of support vector machines for optimal operation rules of Zayandehrood dam

Optimal reservoir operation is generally a complex problem due to the wide range of important influencing factors. Crop mix pattern optimization and limitations, economic indices of agriculture, industry and water supplier and reservoir operation requirements are examples of the factors that might be taken into account in developing a comprehensive optimal model for reservoir operation. The considerable number of involved variables and parameters as well as their complex interactions increase the complexity of the problem and therefore poses limitations for real-time applications. The management process can be simplified for real-time applications using accurate data mining models. While the reservoir optimum operation model is available a large set of data can be generated with different data sets to represent a wide range of conditions which reservoir may operate under it, including the best decisions for water release and storage. Support Vector Machine (SVM) is a relatively new and promising supervised machine learning technique based on the statistical learning theory which is receiving increasing attention lately in the field of water resources management. The combination of SVM , as a well-known data mining method in acquiring complex patterns of behavior and complexity of real reservoir operation models, and a real reservoir operation model seems promising in presenting a simplified model for reservoir operation while considering a large amount of parameters affecting the reservoir operation.
Herein, a complex reservoir operation model for Zayandehrood dam considering upstream and downstream rainfall-runoff, groundwater supply, crop mix, income and employment in agriculture and industry sectors, water demand of domestic, agriculture and industry sectors and interactions of these different factors is utilized to produce a large set of hydrologic, socio-economic and reservoir operation data based on a long-term optimization approach. In this paper, results of a developed integrated hydrologic-socio-economic optimum water allocation model in the Zayandehrood water basin are used by SVM to derive the optimum operating rules for the Zayandehrood dam. The trained SVM predicts the optimum water release from Zayandehrood reservoir based on upstream and downstream hydrologic and operational indices including monthly precipitation, reservoir initial storage, irrigation, industry and domestic water demands. Zayandehrood reservoir operation model produces the optimal crop mix pattern and level of industry production considering the net profit and employment indices as well as the water authority and domestic water supply profits at upstream and downstream of the Zayandehrood dam's reservoir. Considering this optimal set of the data, optimal value of the water demand according to each set of reservoir storage state and precipitation is determined. The reservoir initial storage volume, upstream and downstream precipitation are then used as inputs to produce the optimal water allocation and reservoir release which are automatically producing the optimal net profit and employment in agriculture, industry, water supply and water authority sectors. Therefore, here optimal reservoir operation data sets are produced that are hydrologically and socio-economically optimal for a long period of 20 years. These data sets contain optimal management decisions of the Zayandehrood dam's reservoir system within the Zayandehrood water basin.
SVM is trained and tested by randomly splitting the 20 year's monthly data analysis results (240 data point for each parameter). The train and test data sets were normalized in the range of 0 to 1 and learning parameters were chosen through an optimization procedure. Considering the fact that choosing the most appropriate kernel is depends on the problem being considered, the key stage in SVM is choosing the appropriate kernel function. Linear, Polynomial, Gaussian and Hyperbolic kernels are the most popular kernel functions that are used in this work. Root Mean Square Error (RMSE) and Correlation Coefficients (CC) are the statistical measures used for choosing the best Kernel function in a trial-error procedure. Results of water allocation model in Zayandehrood water basin and the performance of the trained SVM are compared with Artificial Neural Network (ANN) which is a well-known classical machine learning algorithm. The performance of SVM-based and ANN-based predictions is evaluated and the normality of errors is studied as well. The test results show that the errors are independent and are normally distributed and both models are efficient in determining the rules for optimal reservoir operation. However, the ANN-based predictions had somewhat higher predictive power than the SVM-based predictions.
Optimal reservoir operation is generally a complex problem due to the wide range of important influencing factors. Crop mix pattern optimization and limitations, economic indices of agriculture, industry and water supplier and reservoir operation requirements are examples of the factors that might be taken into account in developing a comprehensive optimal model for reservoir operation. The considerable number of involved variables and parameters as well as their complex interactions increase the complexity of the problem and therefore poses limitations for real-time applications. The management process can be simplified for real-time applications using accurate data mining models. While the reservoir optimum operation model is available a large set of data can be generated with different data sets to represent a wide range of conditions which reservoir may operate under it, including the best decisions for water release and storage. Support Vector Machine (SVM) is a relatively new and promising supervised machine learning technique based on the statistical learning theory which is receiving increasing attention lately in the field of water resources management. The combination of SVM , as a well-known data mining method in acquiring complex patterns of behavior and complexity of real reservoir operation models, and a real reservoir operation model seems promising in presenting a simplified model for reservoir operation while considering a large amount of parameters affecting the reservoir operation