T3 – Forecast-Informed Reservoir Operations using short-medium-long range forecast products

Short description

In water systems management, accurate forecasts are essential for improving operational efficiency, particularly in multi-purpose systems where trade-offs among multiple objectives are severe. However, the practical use of forecasts is often limited because high-quality data is only available for a few recent years. This data scarcity makes it difficult to train and test Forecast-Informed Reservoir Operations (FIRO) strategies using modern products such as the latest Artificial Intelligence-enhanced forecasts [1]. While these products provide high accuracy for 10-day horizons [2], they lack the long-term historical archives needed to validate reservoir operating policies over many decades. To address this challenge and support the design of FIRO solutions via Reinforcement Learning algorithms, synthetic forecast generation [3] can recreate forecast-consistent datasets for historical periods with only observational data.

The primary objective of this thesis is to develop these synthetic forecasts for Lake Como and evaluate their impact on real-world operations. This enables a rigorous comparison of the operational effectiveness of new AI-enhanced global products (AIFS) against physics-based benchmarks (HRES), locally calibrated short-term forecasts (Progea 3-day), and long-range Subseasonal-to-Seasonal (S2S) information.

PROPOSED ACTIVITIES

Literature review

  • Study the state of the art in Forecast-Informed Reservoir Operations and synthetic forecast generation.
  • Review existing forecast products available for the Lake Como case study [4].

Data collection and synthetic forecast generation

  • Acquisition of ECMWF HRES and AIFS hindcasts and operational data.
  • Analysis of forecast skill for the different products.
  • Development of a multi-decadal synthetic forecast dataset (1946–present) using state-of-the-art methods.

Design of FIRO solutions

  • Design of multi-scale operating policies using the Evolutionary Multi-Objective Direct Policy Search method to identify the optimal integration of short-, medium-, and long-range information.
  • Trade-off analysis between flood risk in Como, irrigation reliability for the Adda districts, and ecological preservation.
  • Evaluation of the added value of forecast information, determining which forecast horizons most effectively improve reservoir operations across the three objectives.

References:

  1. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
  2. ECMWF. (2025). ECMWF’s AI forecasts become operational: The Artificial Intelligence Forecasting System (AIFS). European Centre for Medium-Range Weather Forecasts.
  3. Brodeur, Z. P., Taylor, W., Herman, J. D., & Steinschneider, S. (2025). Synthetic ensemble forecasts: Operations-based evaluation and inter-model comparison for reservoir systems across California. Water Resources Research, 61(6), e2024WR039324. https://doi.org/10.1029/2024WR039324
  4. Zanutto, D., Ficchì, A., Giuliani, M., & Castelletti, A. (2025). Reinforcement learning of multi-timescale forecast information for designing operating policies of multi-purpose reservoirs. Water Resources Research, 61(2), e2023WR036724. https://doi.org/10.1029/2023WR036724

Relevant courses and knowledge

Natural Resources Management 2

Number of students

1

Requisites

The student should be comfortable with data handling and programming skills (Matlab or Python).