Some forecasting models for photovoltaic power production are developed and applied in this thesis work. There are included both physical models (with extraction of circuital parameters of a photovoltaic module from datasheet and by fitting of experimental measured V-I curves) and statistical models based on artificial neural networks (ANN); these two categories of methods have been applied starting from measured environmental conditions and 24 hours weather forecast. Physical models consider the double version of NOCT and SANDIA thermal models (applied for two 3-parameters models, datasheet and fitting, and two 5-parameters models). The period of training for ANN is seven day long, for different configurations: pure ANN (only weather forecast data as input), hybrid ANN (with solar radiation clear sky model in addition to weather forecast data) and ANN trained with forecast power by physical method (in addition to forecast data and radiation from clear sky model, the power predicted by a physical method is given as target for the training instead of the measured power, which is not available for new plants). Analysis of results is performed calculating hourly errors (deviation of predicted hourly power from measured power) and daily errors; among these ones NMAE (normalized mean absolute error) results to be the most significant: it is a mean of the absolute hourly errors during daylight hours normalized on the rated power of a photovoltaic plant. Physical models result to be more accurate when starting from measured environmental conditions, with NMAE errors lower than 5 % in sunny days. In case of prevision starting by weather forecast there is a general increment for all the error indexes, due to the uncertainty which normally affects the weather forecast. In this case methods which have the lowest errors are those based on hybrid ANN (with clear sky model) and ANN trained with forecast power by physical method; this is true except for those days that are very different from the ones used for the training phase, characterized by NMAE maximum errors of about 35 % (versus the maximum of 20 % of physical methods). The worst results of physical methods are registered in presence of days in which the irradiance is very low and variable or, in the case of the weather forecast, when the forecast irradiance is very different from the measured one. On the other hand artificial neural networks have high error if the forecast day is very different from the ones utilized in the training phase.
Nel lavoro di tesi vengono sviluppati e applicati dei modelli previsionali per la potenza prodotta da moduli fotovoltaici. Si considerano sia modelli fisici (con estrazione dei parametri da datasheet e tramite fitting di curve V-I misurate sperimentalmente) sia modelli statistici basati sulle reti neurali artificiali; entrambe le categorie si applicano partendo da dati misurati da una centralina meteo e da dati forniti dalle previsioni meteo a 24 ore. Per i modelli fisici, di cui si applicano due modelli a tre parametri (datasheet e fitting) e due a cinque parametri, si considerano i modelli termici NOCT e SANDIA. Le reti neurali, addestrate con sette giorni di storico dati, si propongono in diverse configurazioni: pure (solo dati da previsioni meteo in input), ibride (con dati da previsioni meteo e radiazione solare da modello clear sky) e ibride con irraggiamento da fisico (oltre ai dati da previsioni meteo e alla radiazione solare da clear sky si fornisce come target la potenza risultante dall’applicazione di un metodo fisico al posto di quella misurata, non disponibile per un nuovo impianto). L’analisi dei risultati è effettuata in termini di errori orari (differenza tra la potenza oraria prevista e quella misurata) e di errori giornalieri; tra questi ultimi significativo è l’NMAE (normalized mean absolute error), ovvero una media giornaliera degli errori orari assoluti effettuata sulle ore di irraggiamento non nullo normalizzata sulla potenza nominale dell’impianto fotovoltaico. Nel caso di previsioni a partire da dati misurati dalla centralina meteo i metodi fisici risultano essere più precisi, con NMAE ottenuti inferiori al 5 % in giorni in cui non si ha nuvolosità troppo variabile. Nel caso di previsioni a partire dai dati forniti dal servizio meteorologico si ha un generale aumento di tutti gli errori dovuto all’incertezza delle previsioni meteo. I metodi che producono previsioni migliori sono quelli basati sulla rete neurale ibrida con clear sky e sulla rete neurale con potenza prevista da metodo fisico; ciò è vero tranne che in alcuni giorni in cui il giorno previsto dalle reti neurali si discosta molto dai giorni di training, per cui si ottengono errori NMAE con picchi del 35 % (contro il 20 % di picco dei fisici). Lo svantaggio dei metodi fisici è la presenza di giornate in cui si abbia un irraggiamento molto basso e variabile oppure, nel caso delle previsioni meteo, un irraggiamento previsto molto diverso da quello misurato. Le reti neurali hanno invece lo svantaggio di fornire errori elevati se il giorno previsto è molto diverso da quelli utilizzati per il training.
Sviluppo di modelli previsionali per il fotovoltaico : metodi fisici e reti neurali artificiali
MICCO, ANGELO
2014/2015
Abstract
Some forecasting models for photovoltaic power production are developed and applied in this thesis work. There are included both physical models (with extraction of circuital parameters of a photovoltaic module from datasheet and by fitting of experimental measured V-I curves) and statistical models based on artificial neural networks (ANN); these two categories of methods have been applied starting from measured environmental conditions and 24 hours weather forecast. Physical models consider the double version of NOCT and SANDIA thermal models (applied for two 3-parameters models, datasheet and fitting, and two 5-parameters models). The period of training for ANN is seven day long, for different configurations: pure ANN (only weather forecast data as input), hybrid ANN (with solar radiation clear sky model in addition to weather forecast data) and ANN trained with forecast power by physical method (in addition to forecast data and radiation from clear sky model, the power predicted by a physical method is given as target for the training instead of the measured power, which is not available for new plants). Analysis of results is performed calculating hourly errors (deviation of predicted hourly power from measured power) and daily errors; among these ones NMAE (normalized mean absolute error) results to be the most significant: it is a mean of the absolute hourly errors during daylight hours normalized on the rated power of a photovoltaic plant. Physical models result to be more accurate when starting from measured environmental conditions, with NMAE errors lower than 5 % in sunny days. In case of prevision starting by weather forecast there is a general increment for all the error indexes, due to the uncertainty which normally affects the weather forecast. In this case methods which have the lowest errors are those based on hybrid ANN (with clear sky model) and ANN trained with forecast power by physical method; this is true except for those days that are very different from the ones used for the training phase, characterized by NMAE maximum errors of about 35 % (versus the maximum of 20 % of physical methods). The worst results of physical methods are registered in presence of days in which the irradiance is very low and variable or, in the case of the weather forecast, when the forecast irradiance is very different from the measured one. On the other hand artificial neural networks have high error if the forecast day is very different from the ones utilized in the training phase.| File | Dimensione | Formato | |
|---|---|---|---|
|
2015_12_Micco.pdf
solo utenti autorizzati dal 03/12/2016
Descrizione: Testo della tesi
Dimensione
7.18 MB
Formato
Adobe PDF
|
7.18 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/10589/116841