Evaluating LSTM Neural Networks for Energy Forecasting in CHP Plants: A Comparative Approach
(2025) DABN01 20251Department of Economics
Department of Statistics
- Abstract
- This paper evaluates the performance of the Long Short-Term Memory (LSTM) Neural Network for forecasting energy production from combined heat and power (CHP) plants in Stuttgart, Germany. The dataset consists of high frequency time series data and exhibits multiple seasonal patterns. A key challenge addressed in this study is the presence of a systematic reporting delay, where the most recent four days of data are not available for forecasting. To handle this, a hybrid approach is adopted, using a LightGBM model to impute the missing data caused by the reporting delay. This model also serves as a benchmark, alongside a dynamic harmonic regression model. The results show that the LSTM model is neither more accurate nor more computationally... (More)
- This paper evaluates the performance of the Long Short-Term Memory (LSTM) Neural Network for forecasting energy production from combined heat and power (CHP) plants in Stuttgart, Germany. The dataset consists of high frequency time series data and exhibits multiple seasonal patterns. A key challenge addressed in this study is the presence of a systematic reporting delay, where the most recent four days of data are not available for forecasting. To handle this, a hybrid approach is adopted, using a LightGBM model to impute the missing data caused by the reporting delay. This model also serves as a benchmark, alongside a dynamic harmonic regression model. The results show that the LSTM model is neither more accurate nor more computationally efficient than the LightGBM model. However, both models outperform the statistical and naive benchmarks. Increasing sequence length for an LSTM model does not necessarily increase forecasting accuracy due to excessively long input sequences with high frequency data. This paper adds to the existing literature regarding advanced machine learning models’ role in energy forecasting. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9199188
- author
- Hult, Felix LU and Ifantidis Gharizadeh, Evagelos LU
- supervisor
-
- Simon Reese LU
- organization
- course
- DABN01 20251
- year
- 2025
- type
- H1 - Master's Degree (One Year)
- subject
- keywords
- energy forecasting, machine learning, LSTM, LightGBM, reporting delay
- language
- English
- id
- 9199188
- date added to LUP
- 2025-09-12 09:04:25
- date last changed
- 2025-09-12 09:04:25
@misc{9199188, abstract = {{This paper evaluates the performance of the Long Short-Term Memory (LSTM) Neural Network for forecasting energy production from combined heat and power (CHP) plants in Stuttgart, Germany. The dataset consists of high frequency time series data and exhibits multiple seasonal patterns. A key challenge addressed in this study is the presence of a systematic reporting delay, where the most recent four days of data are not available for forecasting. To handle this, a hybrid approach is adopted, using a LightGBM model to impute the missing data caused by the reporting delay. This model also serves as a benchmark, alongside a dynamic harmonic regression model. The results show that the LSTM model is neither more accurate nor more computationally efficient than the LightGBM model. However, both models outperform the statistical and naive benchmarks. Increasing sequence length for an LSTM model does not necessarily increase forecasting accuracy due to excessively long input sequences with high frequency data. This paper adds to the existing literature regarding advanced machine learning models’ role in energy forecasting.}}, author = {{Hult, Felix and Ifantidis Gharizadeh, Evagelos}}, language = {{eng}}, note = {{Student Paper}}, title = {{Evaluating LSTM Neural Networks for Energy Forecasting in CHP Plants: A Comparative Approach}}, year = {{2025}}, }