Monthly Reservoir Inflow Prediction by Neural Network in Thailand

Master(2017) Seiya MARUYAMA

 It is important to make appropriate water management for disaster prevention toward flood or drought in Thailand. To achieve better water management, well-ordered reservoir operation in river is needed.

 There is basic water management cycle – storing water in wet season and using water in dry season – in Thailand because weather of Thailand can be divided clearly into dry season (November to April) and wet season (May to October). Due to sustaining water for dry season, dams which have large storage is needed. For example, Bhumibol Dam and Sirikit Dam which located in northern Chao-Phraya River have larger storage than dam in Japan (Figure in below).

• Figures were cited from Landsat satellite data.
• About capacity data: Capacity of Bhumibol Dam and Sirikit Dam were cited from Electricity Generating Authority of Thailand (EGAT) homepage (*1). Capacity of Tokuyama Dam were cited from Dam Management Office of Tokuyama Dam homepage (*2).

 On the other hand, if reservoir inflow would exceed reservoir capacity, operator must discharge water gradually. In that situation, when operator can know near future reservoir inflow in wet season, they can make optimal discharge plan for both flood and drought prevention. Prediction of reservoir inflow toward several month has possibility to play an important role for water management in Thailand.

 Recently, there are various amount of observed data. Therefore, Statistical model applying observed data is effective method. Neural Network is one of the statistical method for conducting prediction. In this research, 1 month toward prediction of reservoir inflow was conducted by Recurrent Neural Network. Recurrent Neural Network is thought to be effective for time series prediction due to structure.

 Figure in below shows the result of prediction by Recurrent Neural Network. Each red line indicates predicted line from simplex parameters. It is difficult to obtain better prediction by only simplex parameter. Then, several Neural Network model was constructed by learning which started with different initial condition. We can get tendency of 1 month toward reservoir inflow by this method.

Example of 1 month toward reservoir inflow prediction in Sirikit Dam

Horizontal axis: Period (Jan. 2011 - Dec. 2014, Monthly)     Vertical axis: Inflow volume (unit: m3)

Black line shows Observed volume and Red lines show Predicted volume.

 This research regarded 1 month forward prediction as first step. In the future, if we can obtain tendency or fluctuation of reservoir inflow toward several month, decision making of reservoir operation will be easier than current situation. One of the final goal of this research is to provide useful idea for water management in Thailand.

Citation:

*1  Electricity Generating Authority of Thailand, Information, Power Plants and Dams

      (https://www.egat.co.th/en/information/power-plants-and-dams)

*2  独立行政法人　水資源機構　徳山ダム管理所，徳山ダムの概要

      (http://www.water.go.jp/chubu/tokuyama/gaiyo/index.html)