In the present thesis, data-driven methodologies, aiming at reducing the energy consumption of a hospital complex's district heating system, are proposed and implemented. For this purpose, feature generation procedures are first applied to the historical measured consumption values and the corresponding climatic condition data in order to generate a comprehensive dataset. Data-driven pipelines are then developed and tuned for hour ahead prediction of the thermal load of each building that is fed by the district heating system. Deep Learning algorithms and in particular Recurrent neural networks (RNNs), which are designed to operate over sequential or time series data, are employed. Long short term memory (LSTM) algorithm is specifically utilized as the most promising alternative; a choice that is motivated by its capability of overcoming the vanishing gradients problem of the standard RNN when dealing with long term dependencies. Different trained models are then compared based on prediction accuracy, complexity and calculation cost. Optimal prediction pipeline including the determined most promising tuning parameters is then determined leading to an elevated forecasting accuracy. The predicted thermal loads are then fed to the implemented data-driven models of the buildings' heat exchangers, which can thus estimate the resulting primary water's return temperature for any given supply water temperature. Employing these models, an optimization procedure is run in hourly manner which can determine the optimal supply temperature which minimizes the overall return temperature. The resulting reduction in the average temperature of primary water in the network, compared to commonly used outside temperature based control strategy, lead to a notable reduction in the distribution losses. The project is conducted in the context of a collaboration with Siram - by Veolia SpA (The energy management firm of the considered case study) and Concordia Institute for Information Systems Engineering.
Nella presente tesi vengono proposte e implementate metodologie basate su datiper ridurre il consumo energetico del sistema di teleriscaldamento di un complessoospedaliero. A questo scopo, vengono dapprima applicate procedure di feature gen-eration sui valori storici dei consumi misurati e sui corrispondenti dati climatici,al fine di generare un dataset completo. Le pipeline guidate da dati vengono poisviluppate e messe a punto per previsione del carico termico di ogni edificio alimen-tato dal sistema di teleriscaldamento. Vengono utilizzati algoritmi di Deep Learninge in particolare le reti neurali ricorrenti (RNN), che sono progettate per funzionaresu dati sequenziali o su serie temporali. L’algoritmo della memoria a breve terminea lungo termine (LSTM) `e specificamente utilizzato come l’alternativa pi`u promet-tente; una scelta motivata dalla sua capacit`a di superare il problema della scomparsadei gradienti dello standard RNNs quando si tratta di dipendenze a lungo termine.Diversi modelli addestrati, vengono poi confrontati in base all’accuratezza delle pre-visioni, alla complessit`a e ai costi di calcolo. Vengono quindi determinate le pipelinedi previsione ottimale, compresi i parametri pi`u promettenti ottenuti, il che portaad un’elevata precisione di previsione.I carichi termici previsti vengono poi alimentati ai modelli implementati basatisu dati degli scambiatori di calore degli edifici, che possono cos`ı stimare la temper-atura di ritorno dell’acqua primaria risultante per qualsiasi temperatura dell’acquadi alimentazione. Utilizzando questi modelli, viene eseguita una procedura di ot-timizzazione su base oraria in grado di determinare la temperatura ottimale di man-data che riduce al minimo la temperatura di ritorno complessiva. La conseguenteriduzione della temperatura media dell’acqua primaria della rete, rispetto alle co-muni strategie di controllo basate sulla temperatura esterna, porta ad una notevoleriduzione delle perdite di distribuzione. Il progetto `e stato condotto nell’ambito diuna collaborazione con Siram - By Veolia SpA (la societ`a di gestione energetica delcaso studio considerato) e Concordia Institute for Information Systems Engineering.
Application of deep learning in thermal load forecasting and data-driven supply optimization of a district heating network
BOLOURCHIFARD, FARSHAD
2018/2019
Abstract
In the present thesis, data-driven methodologies, aiming at reducing the energy consumption of a hospital complex's district heating system, are proposed and implemented. For this purpose, feature generation procedures are first applied to the historical measured consumption values and the corresponding climatic condition data in order to generate a comprehensive dataset. Data-driven pipelines are then developed and tuned for hour ahead prediction of the thermal load of each building that is fed by the district heating system. Deep Learning algorithms and in particular Recurrent neural networks (RNNs), which are designed to operate over sequential or time series data, are employed. Long short term memory (LSTM) algorithm is specifically utilized as the most promising alternative; a choice that is motivated by its capability of overcoming the vanishing gradients problem of the standard RNN when dealing with long term dependencies. Different trained models are then compared based on prediction accuracy, complexity and calculation cost. Optimal prediction pipeline including the determined most promising tuning parameters is then determined leading to an elevated forecasting accuracy. The predicted thermal loads are then fed to the implemented data-driven models of the buildings' heat exchangers, which can thus estimate the resulting primary water's return temperature for any given supply water temperature. Employing these models, an optimization procedure is run in hourly manner which can determine the optimal supply temperature which minimizes the overall return temperature. The resulting reduction in the average temperature of primary water in the network, compared to commonly used outside temperature based control strategy, lead to a notable reduction in the distribution losses. The project is conducted in the context of a collaboration with Siram - by Veolia SpA (The energy management firm of the considered case study) and Concordia Institute for Information Systems Engineering.File | Dimensione | Formato | |
---|---|---|---|
Thesis_Farshad.pdf
non accessibile
Descrizione: Thesis text
Dimensione
4.02 MB
Formato
Adobe PDF
|
4.02 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/148823