The intelligent decision-making process for hydro scheduling

At present, no matter how sophisticated the optimization tools are, the hydropower operators have to manually set up the executive commands before running the optimization models. Though various solution methods or heuristics have been developed and are available for use, limited by human analytic competence, the operators prefer to choose the commands they are familiar with or adopt the default setting directly. These commands are only set up once before optimization and are valid for all the hydraulic objects (i.e., reservoirs, plants, hydro-turbine generator units) and the entire scheduling horizon. The innovation of this project is to replace the current manual setup of commands with the automatic allocation of commands depending on the specific operating and market conditions of the given hydro system. This automation will be achieved by integrating modern machine learning techniques with a comprehensive understanding of the hydro systems and optimization models. This project pioneers a brand new decision-making process for hydro scheduling. It stretches out the application of the optimization tools to a broader decision-making level, from a uniform scheduling tool to a customized intelligent scheduling system. The methodology developed in the project can be applied to a wider context, not only limited to hydro scheduling, since almost all the advanced decision support tools designed for large-scale problems confront such a similar challenge as the pre-selection of suitable executive commands. During the entire duration of the project period, 1) All the participating industrial partners identify the real-world hydro scheduling problems and provide historical operating data of operating and market conditions. 2) SINTEF determines the corresponding commands to solve the specific problems and generates the datasets. A dataset includes the historical input data for a given hydro system, i.e., inflow, market price, initial water level, and end water value for each reservoir. It also includes the command setting that gives the best solution, i.e., higher objective value, lower calculation time, or less nonphysical spill. 3) NTNU looks into the datasets and the pre-processing techniques to reduce the number of the features, i.e., the number of input columns. Different supervised learning models, such as Random Forest Classifier, k-Nearest Neighbors, Multi-layer Perceptron, Support Vector Machines, Naive Bayes, AdaBoost Classifier, and Histogram based gradient boosting, are tested and a balanced accuracy was used to find the correct machine learning model for each dataset. 4) SINTEF runs the scheduling tool with the default command setting and the command setting predicted by machine learning, respectively. It is demonstrated that the performance can be improved 25% - 100% by using the command predicted by machine learning. It means that in some cases, nonphysical spills from reservoirs can be largely avoided or calculation time can be significantly reduced. 5) SINTEF develops an interactive Jupyter notebook where all the industrial partners can easily test the machine learning models and compare the results for their own hydro systems. This Jupyter notebook can be a solid foundation for further developing other intelligent decision-making processes.

The achieved outcomes of this project are to prove that machine learning techniques can be applied in the decision-making process for hydro scheduling. The following increased value creation is realized: 1) the increased system awareness and improved decision quality provided by the advanced analysis methods; 2) the extra profit gained from the tailored determination of solution methods; 3) the significant reduction in calculational time; 4) the successful prevention of nonphysical spills from reservoirs. One scientific paper has been published and two working papers and one master thesis are in progress. An interactive Jupyter notebook has been developed. All the industrial partners can easily test the machine learning models and compare the results for their own hydro systems. This Jupyter notebook can be a solid foundation for further developing other intelligent decision-making processes.

In this project, Norwegian hydropower producers will take a significant step towards the goal of a wholly automated hydro scheduling system. Instead of improving or replacing any existing optimization tools, the attention will be paid to the way that these tools are used for advanced decision support. This project will focus on the pre-setting of the executive commands that are called to run the optimization tools. At present, no matter how sophisticated the optimization tools are, the hydropower operators have to manually set up the executive commands before running the optimization models. Though various solution methods or heuristics have been developed and are free to use, limited by human analytic competence, the operators prefer to choose the commands they are familiar with or directly adopt the default setting. These commands are only set up once before optimization and valid for all the hydraulic objects (i.e., reservoirs, plants, hydro-turbine generator units) and the entire scheduling horizon. The planned innovation of this project is to replace the current manual setup of commands with the automatic allocation of commands depending on the specific operating and market conditions of the given hydro system. This automation will be achieved by integrating modern machine learning techniques with a comprehensive understanding of the hydro systems and optimization models. This project pioneers a brand new decision-making process for hydro scheduling. It stretches out the application of the optimization tools to a broader decision-making level, from a uniform scheduling tool to a customized intelligent scheduling system. If successful, the methodology developed in the project can be applied to a wider context, not only limited to hydro scheduling, since almost all the advanced decision support tools designed for large-scale problems confront such a similar challenge as the pre-selection of suitable executive commands.