Today, the Local Public Transport sector is required to comply with increasingly rigorous service quality reporting standards. In Italy, Resolution No. 53/2024 of the Transport Regulation Authority has recently mandated the calculation and monitoring of standardized key performance indicators. However, many companies still lack structured tools to transform raw operational data from automatic systems into usable insights for regulatory compliance and decision support. This work, born as part of a curricular internship at Pluservice S.r.l., addresses this need by designing and developing a software platform structured into two complementary components. A Business Intelligence system, developed in Power BI, is dedicated to monitoring the indicators prescribed by the resolution through interactive dashboards, complemented by supporting analyses across various operational dimensions of the service. Additionally, a forecasting system is implemented to predict vehicle delays at scheduled stops. On this front, the study aims to answer two core research questions: first, evaluating the effectiveness of time-series forecasting techniques applied to transit data by comparing classical statistical methods (SARIMAX, Prophet) with neural network approaches (NeuralProphet, LSTM); second, assessing whether a single global model, trained on all available time series, can compete with a local approach. The methodology integrates the analysis and reorganization of operational data, the design of a unified data model for the Business Intelligence dashboards, and an experimental forecasting framework structured at two different granularities (a line-level analysis with a 14-day horizon, and a single-stop with one-step-ahead horizon). These setups compare the different models under a standardized evaluation protocol. The empirical results demonstrate that performances depend strictly on the granularity and forecasting horizon, with SARIMAX proving more accurate over short horizons, and LSTM delivering superior results at the stop level and over longer horizons. Furthermore, the global approach proves advantageous in both tasks, showing greater performance gains as the number of series available for feature extraction and training increases.
Il settore del Trasporto Pubblico Locale è chiamato oggi a una rendicontazione sempre più rigorosa della qualità del servizio. In Italia, la delibera n.~53/2024 dell'Autorità di Regolazione dei Trasporti ha recentemente imposto il calcolo di indicatori standardizzati di performance. Molte aziende, tuttavia, non dispongono ancora di strumenti strutturati per trasformare i dati operativi grezzi dei sistemi di localizzazione automatica dei veicoli in informazioni utili al monitoraggio normativo e al supporto decisionale. Il presente lavoro, svolto nell'ambito di un tirocinio curricolare presso la Pluservice S.r.l., affronta questa esigenza attraverso la progettazione e lo sviluppo di una piattaforma software, articolata in due componenti complementari. Da un lato un sistema di Business Intelligence, realizzato in Power BI, dedicato al monitoraggio degli indicatori previsti dalla delibera attraverso dashboard interattive, integrate da analisi di supporto sulle diverse dimensioni operative del servizio. Dall'altro, un sistema di forecasting dedicato alla previsione del ritardo registrato alle fermate del servizio. Su quest'ultimo fronte, il lavoro si propone di rispondere a due domande di ricerca: quale sia l'efficacia delle principali tecniche di previsione su serie temporali di trasporto, confrontando approcci statistici classici (SARIMAX, Prophet) con approcci neurali (NeuralProphet, LSTM); e se un singolo modello globale, addestrato su tutte le serie disponibili, sia in grado di competere con un approccio locale, basato su un modello per ciascuna serie. La metodologia combina l'analisi e la riorganizzazione dei dati operativi disponibili, la progettazione di un modello dati condiviso per le dashboard di Business Intelligence, e un impianto sperimentale di forecasting strutturato su due granularità (a livello di linea, con orizzonte di 14 giorni, e di singola fermata, a un passo temporale) confrontando diversi modelli secondo un protocollo di valutazione comune. I risultati dimostrano che l'efficacia dei modelli dipende dalla granularità e dall'orizzonte considerati, con SARIMAX più accurato su orizzonte breve e il LSTM superiore a livello di fermata e orizzonti lunghi. Infine, l'approccio globale si conferma vantaggioso in entrambi i task, con performance migliori all'aumentare delle serie da cui estrarre la previsione.
Supporto decisionale data-driven per il trasporto pubblico locale: progettazione e sviluppo di una piattaforma di Business Intelligence
BICCHERI, EMANUELE
2025/2026
Abstract
Today, the Local Public Transport sector is required to comply with increasingly rigorous service quality reporting standards. In Italy, Resolution No. 53/2024 of the Transport Regulation Authority has recently mandated the calculation and monitoring of standardized key performance indicators. However, many companies still lack structured tools to transform raw operational data from automatic systems into usable insights for regulatory compliance and decision support. This work, born as part of a curricular internship at Pluservice S.r.l., addresses this need by designing and developing a software platform structured into two complementary components. A Business Intelligence system, developed in Power BI, is dedicated to monitoring the indicators prescribed by the resolution through interactive dashboards, complemented by supporting analyses across various operational dimensions of the service. Additionally, a forecasting system is implemented to predict vehicle delays at scheduled stops. On this front, the study aims to answer two core research questions: first, evaluating the effectiveness of time-series forecasting techniques applied to transit data by comparing classical statistical methods (SARIMAX, Prophet) with neural network approaches (NeuralProphet, LSTM); second, assessing whether a single global model, trained on all available time series, can compete with a local approach. The methodology integrates the analysis and reorganization of operational data, the design of a unified data model for the Business Intelligence dashboards, and an experimental forecasting framework structured at two different granularities (a line-level analysis with a 14-day horizon, and a single-stop with one-step-ahead horizon). These setups compare the different models under a standardized evaluation protocol. The empirical results demonstrate that performances depend strictly on the granularity and forecasting horizon, with SARIMAX proving more accurate over short horizons, and LSTM delivering superior results at the stop level and over longer horizons. Furthermore, the global approach proves advantageous in both tasks, showing greater performance gains as the number of series available for feature extraction and training increases.| File | Dimensione | Formato | |
|---|---|---|---|
|
tesi.pdf
embargo fino al 11/01/2028
Dimensione
3.57 MB
Formato
Adobe PDF
|
3.57 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/27779