Modelling Historical Reservoir Releases with Recurrent Deep Learning Methods
(2026) In 2026:E78 BERM02 20261Mathematics (Faculty of Sciences)
- Abstract
- Reliable representation of reservoir operations in large-scale hydrological models remains challenging because reservoir releases are shaped not only by highly variable hydrological conditions but also by complex human operating objectives. While simplified rule-based schemes are commonly used, they often struggle to reproduce observed release behaviour across diverse reservoirs.This thesis evaluates whether recurrent machine learning models can reproduce historical daily reservoir outflows across multiple reservoirs using recentdaily inflow, storage, and cyclical day-of-year information as dynamic inputs. It further assesses whether conditioning these models on static reservoir attributes, namely main use, broad climate group, and storage... (More)
- Reliable representation of reservoir operations in large-scale hydrological models remains challenging because reservoir releases are shaped not only by highly variable hydrological conditions but also by complex human operating objectives. While simplified rule-based schemes are commonly used, they often struggle to reproduce observed release behaviour across diverse reservoirs.This thesis evaluates whether recurrent machine learning models can reproduce historical daily reservoir outflows across multiple reservoirs using recentdaily inflow, storage, and cyclical day-of-year information as dynamic inputs. It further assesses whether conditioning these models on static reservoir attributes, namely main use, broad climate group, and storage capacity, improves their performance and generalisation.
Long short-term memory (LSTM) and gated recurrent unit (GRU) models were trained to simulate historical reservoir outflow using these dynamic inputs. Both standard and conditioned versions of each architecture were evaluated. The conditioned models used the static reservoir attributes to initialise the recurrent state. Two dataset versions were used: a main dataset based on ResOpsUS, and an extra dataset that additionally included Chinese reservoir records. For evaluation, 20 reservoirs were held out entirely to test generalisation to unseen reservoirs, while the remaining reservoirs were split temporally into training, validation, and in-sample test periods. Models were trained by minimising MAE on normalised outflow, and hyperparameters were selected using normalised validation MAE. Model performance was assessed using MAE, RMSE, Nash–Sutcliffe efficiency, and Kling–Gupta efficiency. Additional experiments tested the effect of lookback-window length, with seven-, three- and one-day windows being investigated, and assessed the models’ robustness to Gaussian noise.
The results show that recurrent models are suitable for the task examined in this thesis. Conditioned LSTM models achieved the strongest validation and in-sample test performance, suggesting that static reservoir attributes help when predicting later periods for reservoirs represented during training. However, this improvement did not transfer consistently to fully held-out reservoirs, indicating that generalisation to unseen reservoirs remains challenging. The seven-day lookback window gave the best overall performance, while robustness testing showed that perturbations to storage caused the largest reduction in performance. Compared with the rule-based baseline, the recurrent models better reproduced the timing and variability of daily releases.
Overall, this thesis shows that recurrent machine learning models are a promising approach for reproducing historical reservoir outflows. Static conditioning can improve temporal generalisation, but further work is needed to improve transferability to fully held-out reservoirs. (Less) - Popular Abstract
- Reservoirs store water for uses like drinking water, farming and generating electricity. For these uses, they decide when and how much water to release. Rules for these releases are usually not available publicly, so can be difficult to predict in large water models.
This thesis tests whether computer models can learn how reservoirs are operated from historical data. They are given recent information about how much water is entering the reservoir, how much water is already there, and the time of year. Some models are also additionally given simple information about the reservoirs, like its main purpose, the climate around it and its size.
We found that the computer models were better at this task than other methods that are... (More) - Reservoirs store water for uses like drinking water, farming and generating electricity. For these uses, they decide when and how much water to release. Rules for these releases are usually not available publicly, so can be difficult to predict in large water models.
This thesis tests whether computer models can learn how reservoirs are operated from historical data. They are given recent information about how much water is entering the reservoir, how much water is already there, and the time of year. Some models are also additionally given simple information about the reservoirs, like its main purpose, the climate around it and its size.
We found that the computer models were better at this task than other methods that are traditionally used. The models that were given the additional information had the best results for reservoirs that they had already seen in training. However, predicting releases for completely new reservoirs was still difficult, and the simpler models without the additional information still did better on them. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9246066
- author
- Debiere, Mattéo Adrien LU
- supervisor
-
- Zheng Duan LU
- Shaokun He LU
- organization
- course
- BERM02 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- reservoir operation modelling, reservoir releases, recurrent neural networks, LSTM, GRU, conditioned recurrent models, static reservoir attributes, hydrological modelling, machine learning
- publication/series
- 2026:E78
- report number
- LUNFBV-3020-2026
- ISSN
- 1404-6342
- other publication id
- 2026:E78
- language
- English
- id
- 9246066
- date added to LUP
- 2026-09-14 16:16:15
- date last changed
- 2026-09-14 16:16:15
@misc{9246066,
abstract = {{Reliable representation of reservoir operations in large-scale hydrological models remains challenging because reservoir releases are shaped not only by highly variable hydrological conditions but also by complex human operating objectives. While simplified rule-based schemes are commonly used, they often struggle to reproduce observed release behaviour across diverse reservoirs.This thesis evaluates whether recurrent machine learning models can reproduce historical daily reservoir outflows across multiple reservoirs using recentdaily inflow, storage, and cyclical day-of-year information as dynamic inputs. It further assesses whether conditioning these models on static reservoir attributes, namely main use, broad climate group, and storage capacity, improves their performance and generalisation.
Long short-term memory (LSTM) and gated recurrent unit (GRU) models were trained to simulate historical reservoir outflow using these dynamic inputs. Both standard and conditioned versions of each architecture were evaluated. The conditioned models used the static reservoir attributes to initialise the recurrent state. Two dataset versions were used: a main dataset based on ResOpsUS, and an extra dataset that additionally included Chinese reservoir records. For evaluation, 20 reservoirs were held out entirely to test generalisation to unseen reservoirs, while the remaining reservoirs were split temporally into training, validation, and in-sample test periods. Models were trained by minimising MAE on normalised outflow, and hyperparameters were selected using normalised validation MAE. Model performance was assessed using MAE, RMSE, Nash–Sutcliffe efficiency, and Kling–Gupta efficiency. Additional experiments tested the effect of lookback-window length, with seven-, three- and one-day windows being investigated, and assessed the models’ robustness to Gaussian noise.
The results show that recurrent models are suitable for the task examined in this thesis. Conditioned LSTM models achieved the strongest validation and in-sample test performance, suggesting that static reservoir attributes help when predicting later periods for reservoirs represented during training. However, this improvement did not transfer consistently to fully held-out reservoirs, indicating that generalisation to unseen reservoirs remains challenging. The seven-day lookback window gave the best overall performance, while robustness testing showed that perturbations to storage caused the largest reduction in performance. Compared with the rule-based baseline, the recurrent models better reproduced the timing and variability of daily releases.
Overall, this thesis shows that recurrent machine learning models are a promising approach for reproducing historical reservoir outflows. Static conditioning can improve temporal generalisation, but further work is needed to improve transferability to fully held-out reservoirs.}},
author = {{Debiere, Mattéo Adrien}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{2026:E78}},
title = {{Modelling Historical Reservoir Releases with Recurrent Deep Learning Methods}},
year = {{2026}},
}