A robust energy flow predictor based on CNN-LSTM for prosumer-oriented microgrids considering changes in biogas generation

• Published in: Energy (Oxford) Vol. 326; p. 136050
• Main Authors: Maślak, Grzegorz, Orłowski, Przemysław
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2025
• Subjects: biogas; Biogas energy; CNN-LSTM; energy flow; Energy management system; Hyper-parameter tuning; Microgrid; neural networks; prediction; Robust optimisation; Biogas energy; Microgrid; Hyper-parameter tuning; CNN-LSTM; Energy management system; Robust optimisation
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2025.136050

Information, such as reliable demand and generation forecasts is crucial for appropriate energy management in microgrids. However, the most valuable data is the aggregate energy flow in the managed network, rather than its components. The large share of renewable energy in such installations is associated with fluctuations. A promising renewable source that helps counteract them is biogas. However, a source subject to uncertainties which should be taken into account. The work focuses on developing a method for obtaining a CNN-LSTM microgrid energy flow predictor dedicated to energy management systems. The selection of the hyper-parameters of the artificial neural network is formulated as a robust meta-optimisation problem and solved using surrogate optimisation. The solution ensures obtaining a CNN-LSTM network providing the lowest prediction error with the largest change in the share of biogas in the microgrid. Detailed analyses have shown that in addition to obtaining high robustness to biogas, the solution does not lose accuracy while changing the number of prosumers with photovoltaic installations. Beyond meta-optimisation, ranges of certain hyper-parameters providing adequate prediction quality for all analysed prediction horizons are identified. •The solution obtained considers fluctuations in renewable energy generation.•The predictor is robust to changes in the share of biogas in the energy mix.•Optimised hyper-parameter values ensure prediction error values close to minimal.•Solutions can be generalised to different lengths of prediction horizons.