Direct data-driven predictive control for water storage reservoirs

Water storage reservoirs are critical infrastructure for energy production (hydropower), water supply, and flood protection. The uncertainty of meteorological inputs is a major challenge to operating reservoirs to meet these (often conflicting) objectives. However, meteorological forecast skill has improved dramatically in recent decades, and forecast-informed reservoir operation (FIRO) is a promising approach to maximize reservoir benefits and avoid tradeoffs between conflicting objectives. The state-of-the-art for FIRO is a model predictive control (MPC) approach (see Figure 1, pdf attached): (1) forecasted meteorology is fed into a hydrological model that generates a forecasted river inflow to the reservoir; (2) reservoir operations are optimized for the forecasted inflow; (3) the optimized operations are applied for one stage (e.g., one day); and (4) a new meteorological forecast is received and the process is repeated. Yet, the hydrological processes that transform meteorology into river flow are complex and uncertain themselves. The two-step process of optimizing hydrological model parameters to historical observations and then optimizing reservoir operations to model-forecasted river flow may inadvertently yield suboptimal operating policies. This project proposes an alternative, data-driven approach—rather than attempting to model the complex dynamics between weather forecasts/measurements and reservoir river inflow, the data-driven approach will learn these dynamics from data (see Figure 2, pdf attached). That is, a Data-Driven Predictive Controller (DDPC) will use the learned dynamics to make reservoir release decisions. Specifically, the proposed DDPC controller will use a Transient-Predictor DDPC. The Transient Predictor, recently proposed by Institute for Automatic Control (IFA), improves upon prior DDPC methods by making causal multi-step predictions—inputs at timestep k do not affect outputs prior to time k. In addition to the Transient Predictor, proposed DDPC controller will also include an optimal regularization term that has been derived in previous work. This optimal regularization term allows the DDPC to account for system uncertainty due to the noise in the data, favoring control actions that align with the true system dynamics. DDPC reservoir operation is a novel concept and Transient Predictor DDPC is a new control method. Thus, this thesis seeks to make a notable contribution to data-driven reservoir management. A successful thesis may be converted into a publication in a high impact conference or journal. (see pdf for further details)

Water storage reservoirs are critical infrastructure for energy production (hydropower), water supply, and flood protection. The uncertainty of meteorological inputs is a major challenge to operating reservoirs to meet these (often conflicting) objectives. However, meteorological forecast skill has improved dramatically in recent decades, and forecast-informed reservoir operation (FIRO) is a promising approach to maximize reservoir benefits and avoid tradeoffs between conflicting objectives. The state-of-the-art for FIRO is a model predictive control (MPC) approach (see Figure 1, pdf attached): (1) forecasted meteorology is fed into a hydrological model that generates a forecasted river inflow to the reservoir; (2) reservoir operations are optimized for the forecasted inflow; (3) the optimized operations are applied for one stage (e.g., one day); and (4) a new meteorological forecast is received and the process is repeated. Yet, the hydrological processes that transform meteorology into river flow are complex and uncertain themselves. The two-step process of optimizing hydrological model parameters to historical observations and then optimizing reservoir operations to model-forecasted river flow may inadvertently yield suboptimal operating policies.

This project proposes an alternative, data-driven approach—rather than attempting to model the complex dynamics between weather forecasts/measurements and reservoir river inflow, the data-driven approach will learn these dynamics from data (see Figure 2, pdf attached). That is, a Data-Driven Predictive Controller (DDPC) will use the learned dynamics to make reservoir release decisions.

Specifically, the proposed DDPC controller will use a Transient-Predictor DDPC. The Transient Predictor, recently proposed by Institute for Automatic Control (IFA), improves upon prior DDPC methods by making causal multi-step predictions—inputs at timestep k do not affect outputs prior to time k. In addition to the Transient Predictor, proposed DDPC controller will also include an optimal regularization term that has been derived in previous work. This optimal regularization term allows the DDPC to account for system uncertainty due to the noise in the data, favoring control actions that align with the true system dynamics.

DDPC reservoir operation is a novel concept and Transient Predictor DDPC is a new control method. Thus, this thesis seeks to make a notable contribution to data-driven reservoir management. A successful thesis may be converted into a publication in a high impact conference or journal.

(see pdf for further details)

The primary objectives of the project are to: • Develop a reproducible workflow for applying DDPC to reservoir operation (considering the particularities of the hydrological context). • Demonstrate the application of DDPC for a testbed reservoir. • Identify opportunities and obstacles for further application of DDPC to reservoir operation.

The primary objectives of the project are to:
• Develop a reproducible workflow for applying DDPC to reservoir operation (considering the particularities of the hydrological context).
• Demonstrate the application of DDPC for a testbed reservoir.
• Identify opportunities and obstacles for further application of DDPC to reservoir operation.

Please send your resume/CV, transcript of records, and a brief statement of your interest in this project via email to Kevin Wallington (kwallington@ethz.ch) and Keith Moffat (kmoffat@ethz.ch). Competitive candidates will demonstrate multiple (but not necessarily all) of the following attributes: - Experience in the field of control systems or related fields - especially predictive control or system identification - Prior research or work experience that demonstrate the ability to learn and problem-solve independently - Programming skills - Written and verbal communication skills

Please send your resume/CV, transcript of records, and a brief statement of your interest in this project via email to Kevin Wallington (kwallington@ethz.ch) and Keith Moffat (kmoffat@ethz.ch).

Competitive candidates will demonstrate multiple (but not necessarily all) of the following attributes:
- Experience in the field of control systems or related fields - especially predictive control or system identification
- Prior research or work experience that demonstrate the ability to learn and problem-solve independently
- Programming skills
- Written and verbal communication skills