Implementing LSTM machine learning in long-term hydropower scheduling - Master’s thesis

Aug 5, 2026

Upload by Ezra

Energy demand is rising globally while environmental pressures drive increased substitution of fossil fuels with renewable sources. Many new renewables lack the flexibility and reliability needed to stabilize the power grid. Hydropower scheduling can provide grid stability, but its computational complexity makes frequent scheduling difficult. This work evaluates whether machine learning can reduce time complexity, enabling higher-frequency schedules. An LSTM approach is tested after a literature review and shows a 99.7% reduction in computational time, while accuracy using MAPE as low as 67.5% limits replacing it as a standalone solution.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages64

Reading Time2 h 41 min

AccessFree to download

1 Introduction

2 Theory

2.1 Hydropower Plant

2.2 Benders Decomposition

2.3 Machine Learning and Neural Networks

2.4 Syntetic data

3 Literature review

4 Mathematical model

5 Case study

6 Result and discusion

7 Conclution

8 Future work

References

Master’s thesis NT NU Norwegian Un iversity of Science and Technology Faculty of Information Techno logy and Electrical Engineering Department of Electric Power Engineering Øystein Steinsvik Evjen Implementing LSTM machine learning in long-term hydropower scheduling Master’s thesis in Energi og miljø Supervisor: Jayaprakash Rajasekharan Co-supervisor: Jinghao Wang July 2023 Øystein Steinsvik Evjen Implementing LSTM machine learning in long-term hydropower scheduling Master’s thesis in Energi og miljø Supervisor: Jayaprakash Rajasekharan Co-supervisor: Jinghao Wang July 2023 Norwegian University of Science and Technology Faculty of Information Technology and Electrical Engineering Department of Electric Power Engineering Abstract The energy demand is rapidly increasing worldwide, and with the environmental crisis is more people substituting fossil fuels with renewable energy sources. The new renewable energy sources are not as flexible and reliable as the old fossil power sources which can unstabilize the power grid. Hydropower plants can be scheduled, and therefore stabilize the power grid, but the higher complexity in scheduling makes the computational time too high to conduct a hydropower schedule often enough. Therefore might machine learning help with lowering the time complexity, and thereby let the scheduling be done at a higher frequency. From a literature review have the LSTM neural network model been a possible fit for this research. After testing the model the LSTM model had a 99.7% lower computational time, but the accuracy with MAPE evaluation as low as 67.5% did not hold a standard where the model could be recommended as a substitute, but it can be used as a supplement. Sammendrag Energiforbruket øker raskt globalt, og med miljøkrisen erstatter stadig flere mennesker fossile brensler med fornybare energikilder. De nye fornybare energikildene er ikke like fleksible og p˚alitelige som de gamle fossile energikildene, noe som kanføre til ustabilitet i strømnettet. Vannkraftverk kan planlegges, og dermed stabilisere strømnettet, men den økte kompleksiteten i planleggingen gjør at beregningstidenblir for lang til ˚a gjennomføre en vanlig vannkraftplan ofte nok. Derfor kan maskinlæring bidra til˚a redusere tidskompleksiteten og dermed tillate hyppigere planlegging. Etter en litteraturgjennomgang har LSTM nevrale nettverkmodellen vist seg ˚a være en mulig løsning for denne forskningen. Etter ˚a ha testet modellen viste det seg at LSTM-modellen hadde 99,7% lavere beregningstid, men nøyaktigheten med MAPEevalueringen s˚a lav som 67,5% oppfylte ikke en standard der modellen kunne anbefales som en erstatning, men den kan brukes som et supplement. Preface This master thesis is written for the course TET4900-Elektrisk energi og energisystemer, masteroppgave at NTNU the spring semester of 2023 . I extend my gratitude to my supervisor, Associate Professor Jayaprakash Rajasekharan, for his guidance and constructive feedback throughout my journey. I would also especially like to express my appreciation to my co-supervisor, Ph.D. Candidate Jinghao Wang for his opinion and contributions in providing data, and running data for me. His support has been very helpful on to write this master thesis. Table of Contents List of Figures iv List of Tables v 1 Introduction 1 2 Theory 2 2.1 Hydropower Plant ................................ 2 2.1.1 Hydro power scheduling ......................... 4 2.1.2 Load Shedding .............................. 4 2.2 Benders Decomposition ............................. 5 2.3 Machine Learning and Neural Networks .................... 6 2.3.1 Linear Regression ............................. 7 2.3.2 Neural Network .............................. 7 2.3.3 Long Short-Term Memory (LSTM) ................... 9 2.3.4 Bayesian Optimization .......................... 9 2.3.5 Ensemble Models ............................. 10 2.3.6 Performance Measuring ...

FAQs

Why is hydropower scheduling computationally challenging?

Hydropower scheduling supports grid stability, but increased scheduling complexity leads to high computational time, making frequent planning difficult.

How does LSTM affect computational time in this study?

After testing, the LSTM model achieved a 99.7% lower computational time compared with the baseline approach.

Is the LSTM model accurate enough to replace existing scheduling methods?

Using MAPE evaluation, accuracy down to 67.5% did not meet a standard for recommending the model as a substitute, though it can be used as a supplement.

Aug 5, 2026

Upload by Ezra