Titre :	Multi-target electricity load and price forecasting with LSTM-based architectures and RF
Type de document :	document multimédia
Auteurs :	Amel Chellama, Auteur ; Laradj Chellama, Directeur de thèse
Editeur :	Laghouat : Université Amar Telidji - Département d'informatique
Année de publication :	2026
Importance :	75 p.
Accompagnement :	1 disque optique numérique (CD-ROM)
Note générale :	Option : Artificial intelligence and data science
Langues :	Anglais (eng)
Mots-clés :	Electricity Load  Electricity price  Multi-target forecasting  Deep learning  LSTM  BiLSTM  Random forest
Résumé :	Accurate forecasting of electricity load and price is essential for efficient energy system management, particularly under increasing renewable energy integration and market volatility. However, traditional statistical models often fail to capture the nonlinear dynamics of hourly energy data. This thesis compares three model families Random Forest, LSTM, and Bidirectional LSTM (BiLSTM) for the simultaneous forecasting of electricity load and price across five Spanish cities using hourly data (2015–2018), within a unified experimental framework comprising five configurations per model family (75 runs in total). All models are designed as multi-target architectures that jointly predict both targets from a single combined loss. Results show that LSTMTrial5 achieves the best performance for both targets across all five cities without exception, reaching a mean load R2 of 0.9775 and a mean price R2 of 0.9003, the latter exceeding 0.90 in Bilbao and Valencia. BiLSTMTrial5, while slightly below LSTM in absolute accuracy, exhibits substantially lower variance across hyperparameter configurations (price σ = 0.026 vs. LSTM’s 0.044), making it the more stable choice for deployment. Random Forest achieves competitive load accuracy under its best configuration RFTrial4 (R2 = 0.9654), but remains structurally limited for price forecasting due to its inability to encode temporal ordering. The study confirms that price forecasting is intrinsically harder than load forecasting by 7–9 R2 percentage points, with the largest errors concentrated during price spike episodes driven by market dynamics unobservable from the 24-hour input window.
note de thèses :	Mémoire de master en informatique

Multi-target electricity load and price forecasting with LSTM-based architectures and RF [document multimédia] / Amel Chellama, Auteur ; Laradj Chellama, Directeur de thèse . - Laghouat : Université Amar Telidji - Département d'informatique, 2026 . - 75 p. + 1 disque optique numérique (CD-ROM).
Option : Artificial intelligence and data science
Langues : Anglais (eng)

Mots-clés :	Electricity Load  Electricity price  Multi-target forecasting  Deep learning  LSTM  BiLSTM  Random forest
Résumé :	Accurate forecasting of electricity load and price is essential for efficient energy system management, particularly under increasing renewable energy integration and market volatility. However, traditional statistical models often fail to capture the nonlinear dynamics of hourly energy data. This thesis compares three model families Random Forest, LSTM, and Bidirectional LSTM (BiLSTM) for the simultaneous forecasting of electricity load and price across five Spanish cities using hourly data (2015–2018), within a unified experimental framework comprising five configurations per model family (75 runs in total). All models are designed as multi-target architectures that jointly predict both targets from a single combined loss. Results show that LSTMTrial5 achieves the best performance for both targets across all five cities without exception, reaching a mean load R2 of 0.9775 and a mean price R2 of 0.9003, the latter exceeding 0.90 in Bilbao and Valencia. BiLSTMTrial5, while slightly below LSTM in absolute accuracy, exhibits substantially lower variance across hyperparameter configurations (price σ = 0.026 vs. LSTM’s 0.044), making it the more stable choice for deployment. Random Forest achieves competitive load accuracy under its best configuration RFTrial4 (R2 = 0.9654), but remains structurally limited for price forecasting due to its inability to encode temporal ordering. The study confirms that price forecasting is intrinsically harder than load forecasting by 7–9 R2 percentage points, with the largest errors concentrated during price spike episodes driven by market dynamics unobservable from the 24-hour input window.
note de thèses :	Mémoire de master en informatique