Improving heat demand forecasting with feature reduction in an Encoder–Decoder LSTM model

Accurate short-term heat demand forecasting is essential for the efficient operation of District Heating Networks (DHNs). However, many forecasting models rely on large sets of engineered or externally forecasted variables, introducing redundancy, computational overhead, and reduced generalizability...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Energy reports Ročník 14; s. 5048 - 5060
Hlavní autori: Darbandi, Amin, Brockmann, Gerrid, Kriegel, Martin
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier Ltd 01.12.2025
Predmet:
Artificial Intelligence
Data-driven prediction
Demand forecast
Machine learning
Data-driven prediction
Artificial Intelligence
Demand forecast
Machine learning
ISSN:2352-4847, 2352-4847
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Accurate short-term heat demand forecasting is essential for the efficient operation of District Heating Networks (DHNs). However, many forecasting models rely on large sets of engineered or externally forecasted variables, introducing redundancy, computational overhead, and reduced generalizability. This study proposes an Encoder–Decoder (ED) Long Short-Term Memory (LSTM) architecture emphasizing input feature reduction to enhance robustness and accuracy. Using an open-source smart meter dataset, the model is trained and evaluated across ten systematically constructed input feature sets to quantify the influence of each feature group through a feed-forward selection strategy. Results indicate that dimensionality reduction yields substantial performance gains, with MAE and in CV-RMSE reduced by 57% and 64%, respectively. The inclusion of wind speed improves predictive accuracy, while adding solar irradiance results in a degradation of approximately 5 percentage points. The proposed ED-LSTM outperforms benchmark models, including Support Vector Regression (SVR), Seasonal Autoregressive Integrated Moving Average with Exogenous Regressor (SARIMAX), and a baseline two-layer LSTM, with error reductions ranging from 4.8% to 47.1% across MSE, MAE, RMSE, and MAPE metrics. The statistical significance of model performance differences is confirmed through Friedman and Diebold–Mariano tests, supporting the reliability of the comparative results. [Display omitted] •Encoder–Decoder LSTM forecasts DHN heat demand accurately using reduced inputs.•Feature reduction (15 to 4) boosts robustness, limiting reliance on external forecasts.•Nonparametric Friedman and Diebold–Mariano tests validate the recommended IFS.•Model outperforms SVR, SARIMAX, and shallow LSTM with few features and simple design.
ISSN:2352-4847
2352-4847
DOI:10.1016/j.egyr.2025.11.025