The electrical load forecasting is a fundamental technique for load prediction of consumers for a utility. The accurate load forecasting is crucial to Demand Response (DR) programs in the paradigm of smart grids which allows the consumers to reduce their consumption when the system is under congestion. Several load forecasting models have been applied to get the accurate load prediction based on statistical modelling and artificial intelligence based modelling. Artificial Neural Network (ANN) based techniques have been widely used in the recent years and applied to predict the load with high accuracy to participate in the DR programs for commercial, industrial and residential consumers. This thesis work is focused on the use of two ANN based load forecasting techniques i.e. Feed-Forward Neural Network (FNN) and Echo State Network (ESN) on a data set related to commercial buildings and possible application of load forecasting for DR in commercial sector is discussed. Backpropagation (BP) algorithm is considered for the training of FNN while ESN is trained through Reservoir Computing (RC) methods. The data set is clustered by using K-Means clustering algorithm with respect to three different types of the day which include weekdays with low demand, weekends/holidays and weekdays with high demand and then the FNN and ESN models are implemented. The load forecasting accuracy measures, considered for this work, are Mean Absolute Percentage Error (MAPE), Weighted Mean Absolute Error (WMAE), newly introduced Envelope-Weighted Mean Absolute Error (EMAE), Root Mean Square Error (RMSE) and Normalized Root Mean Square Error (NRMSE). The results of both models are compared based on the load forecasting accuracy measures. MAPE is considered for ranking the results of the models while other error measures are also compared with each other. It is concluded that FNN performs better than the ESN due to the fact that its modelling depends on lesser parameters than ESN and careful computational optimization techniques are required to optimize the parameters of ESN for enhancing the accuracy of the model.
La previsione del carico elettrico è fondamentale per la pianificazione dei consumi elettrici. Una previsione accurata del carico è fondamentale per i programmi di Demand Response (DR) nel paradigma delle Smart Grid, per consentire ai consumatori di ridurre il loro consumo quando il sistema è congestionato. Sono stati analizzati diversi modelli di previsione del carico per ottenere una previsione accurata del carico basata sulla modellizzazione statistica e sull'intelligenza artificiale. Negli ultimi anni le tecniche basate sulle reti neurali artificiali (ANN) sono state ampiamente utilizzate e applicate per prevedere i consumi con elevata precisione al fine di partecipare ai programmi di DR per i carichi commerciali, industriali e domestici. Questo lavoro di tesi si concentra sull'uso di due tecniche di previsione del carico basate su ANN, ad esempio Feed-Forward Neural Network (FNN) ed Echo State Neural Network (ESN) e su un set di dati relativo a edifici commerciali; in particolare, vengono discusse possibili applicazioni di previsione del carico per DR in ambito commerciale. Un algoritmo di Backpropagation (BP) è stato utilizzato per l'addestramento della rete FNN, mentre la rete ESN è stata addestrata attraverso metodi di Reservoir Computing (RC). Il set di dati viene raggruppato utilizzando un algoritmo di clustering K-Means in relazione a tre diverse tipologie di giorni della settimana: con domanda elevata, con bassa domanda, e fine settimana / festività; vengono quindi implementati i modelli FNN ed ESN. Le misure di accuratezza della previsione del carico, considerate per questo lavoro, sono l'errore di percentuale assoluto medio (MAPE), l'errore assoluto medio ponderato (WMAE), l'errore assoluto medio ponderato all’inviluppo (EMAE), l'errore quadratico medio (RMSE) e l'errore quadratico medio normalizzato (NRMSE). I risultati di entrambi i modelli ANN considerati sono confrontati in base alla misura della precisione della previsione del carico. Il MAPE viene primariamente considerato per classificare i risultati dei modelli, mentre le altre misure di errore vengono usate per un confronto addizionale. Si conclude che FNN ha prestazioni migliori rispetto all'ESN, poiché la sua modellazione dipende da un numero di parametri minore rispetto all'ESN, mentre sono necessarie tecniche di ottimizzazione computazionale per ottimizzare i parametri dell'ESN per migliorare l'accuratezza del modello.
Electrical load forecasting using machine learning techniques in view of demand response programs
MANSOOR, MUHAMMAD
2017/2018
Abstract
The electrical load forecasting is a fundamental technique for load prediction of consumers for a utility. The accurate load forecasting is crucial to Demand Response (DR) programs in the paradigm of smart grids which allows the consumers to reduce their consumption when the system is under congestion. Several load forecasting models have been applied to get the accurate load prediction based on statistical modelling and artificial intelligence based modelling. Artificial Neural Network (ANN) based techniques have been widely used in the recent years and applied to predict the load with high accuracy to participate in the DR programs for commercial, industrial and residential consumers. This thesis work is focused on the use of two ANN based load forecasting techniques i.e. Feed-Forward Neural Network (FNN) and Echo State Network (ESN) on a data set related to commercial buildings and possible application of load forecasting for DR in commercial sector is discussed. Backpropagation (BP) algorithm is considered for the training of FNN while ESN is trained through Reservoir Computing (RC) methods. The data set is clustered by using K-Means clustering algorithm with respect to three different types of the day which include weekdays with low demand, weekends/holidays and weekdays with high demand and then the FNN and ESN models are implemented. The load forecasting accuracy measures, considered for this work, are Mean Absolute Percentage Error (MAPE), Weighted Mean Absolute Error (WMAE), newly introduced Envelope-Weighted Mean Absolute Error (EMAE), Root Mean Square Error (RMSE) and Normalized Root Mean Square Error (NRMSE). The results of both models are compared based on the load forecasting accuracy measures. MAPE is considered for ranking the results of the models while other error measures are also compared with each other. It is concluded that FNN performs better than the ESN due to the fact that its modelling depends on lesser parameters than ESN and careful computational optimization techniques are required to optimize the parameters of ESN for enhancing the accuracy of the model.File | Dimensione | Formato | |
---|---|---|---|
2018_07_Mansoor.pdf
non accessibile
Descrizione: Thesis Text
Dimensione
3.44 MB
Formato
Adobe PDF
|
3.44 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/10589/141856