The increasing pressure on energy systems, driven by the convergence of market instability, supply chain vulnerabilities, and the urgency of climate change, has highlighted the limitations of traditional energy models and the need for more efficient management strategies. In this context, the building sector plays a crucial role, particularly buildings equipped with Heating, Ventilation and Air Conditioning (HVAC) systems, which significantly affect both energy consumption and occupants' thermal comfort. This thesis investigates the use of data-driven models based on neural networks as predictive tools to support optimal control of an HVAC system. Two neural network architectures are compared: a conventional Feed-Forward Neural Network (FFNN) and an Input Convex Neural Network (ICNN), both trained on the same dataset to predict the indoor temperature of the building. While the conventional neural network provides maximum predictive flexibility, it does not guarantee any structural properties of the learned function. When such a model is embedded as a constraint within an optimization problem, the non-convexity of the input-output relationship may lead to local minima and hinder the identification of the global optimum. In contrast, the ICNN addresses this limitation by enforcing convexity of the learned function with respect to its inputs, thereby ensuring that the associated optimization problem remains tractable and can be solved reliably. The convex neural network is then integrated into a simplified optimal control formulation, in which the heat pump power, used as a proxy for energy consumption, is minimized while satisfying thermal comfort constraints. Experimental results show that the proposed approach effectively reduces the control input under favorable operating conditions, while also highlighting its limitations in more challenging scenarios.
La crescente pressione sui sistemi energetici, dovuta alla convergenza tra instabilità dei mercati, vulnerabilità degli approvvigionamenti e urgenza climatica, ha evidenziato i limiti dei modelli energetici tradizionali e la necessità di strumenti di gestione più efficienti. In questo contesto, il settore edilizio assume un rilievo particolare, soprattutto per gli edifici dotati di sistemi HVAC (Heating, Ventilation and Air Conditioning), che incidono al tempo stesso sui consumi energetici e sul comfort termico degli occupanti. Questo lavoro di tesi analizza l'impiego di modelli data-driven basati su reti neurali come strumento predittivo e di supporto decisionale per un problema di controllo ottimo applicato a un sistema HVAC. Vengono confrontate due architetture: una rete neurale classica feed-forward (FFNN) e una rete neurale a vincoli di convessità (Input Convex Neural Network, ICNN), entrambe addestrate sullo stesso dataset per la previsione della temperatura interna dell'edificio. Mentre la rete classica offre la massima flessibilità predittiva, non garantisce alcuna proprietà strutturale della funzione appresa: quando un modello di questo tipo viene inserito come vincolo in un problema di ottimizzazione, la non-convessità della relazione ingresso-uscita può generare minimi locali e compromettere la ricerca della soluzione globale. La rete convessa supera questo limite imponendo la convessità della funzione appresa rispetto agli ingressi, una proprietà che rende trattabile in modo affidabile il problema di ottimizzazione associato. La rete convessa viene quindi integrata in una formulazione semplificata di controllo ottimo, in cui la potenza della pompa di calore, assunta come proxy del consumo energetico, viene minimizzata nel rispetto dei vincoli di comfort termico. I risultati sperimentali mostrano che la formulazione proposta consente di ridurre la variabile di controllo nelle condizioni operative favorevoli, mentre ne evidenzia i limiti negli scenari più critici.
MODELLAZIONE DATA-DRIVEN CONVESSA PER IL CONTROLLO OTTIMO DI EDIFICI DOTATI DI SISTEMI HVAC
PALAZZO, SIMONA
2025/2026
Abstract
The increasing pressure on energy systems, driven by the convergence of market instability, supply chain vulnerabilities, and the urgency of climate change, has highlighted the limitations of traditional energy models and the need for more efficient management strategies. In this context, the building sector plays a crucial role, particularly buildings equipped with Heating, Ventilation and Air Conditioning (HVAC) systems, which significantly affect both energy consumption and occupants' thermal comfort. This thesis investigates the use of data-driven models based on neural networks as predictive tools to support optimal control of an HVAC system. Two neural network architectures are compared: a conventional Feed-Forward Neural Network (FFNN) and an Input Convex Neural Network (ICNN), both trained on the same dataset to predict the indoor temperature of the building. While the conventional neural network provides maximum predictive flexibility, it does not guarantee any structural properties of the learned function. When such a model is embedded as a constraint within an optimization problem, the non-convexity of the input-output relationship may lead to local minima and hinder the identification of the global optimum. In contrast, the ICNN addresses this limitation by enforcing convexity of the learned function with respect to its inputs, thereby ensuring that the associated optimization problem remains tractable and can be solved reliably. The convex neural network is then integrated into a simplified optimal control formulation, in which the heat pump power, used as a proxy for energy consumption, is minimized while satisfying thermal comfort constraints. Experimental results show that the proposed approach effectively reduces the control input under favorable operating conditions, while also highlighting its limitations in more challenging scenarios.| File | Dimensione | Formato | |
|---|---|---|---|
|
TESI_PALAZZO_SIMONA.pdf
File consultabile da postazioni biblioteca
Descrizione: Il presente allegato contiene la versione definitiva del lavoro di tesi della sottoscritta Simona Palazzo
Dimensione
4.57 MB
Formato
Adobe PDF
|
4.57 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/65809