The renewable energy transition is able to democratize energy production and unlock vast private capital by embracing distributed energy resources (DERs). These often include solar photovoltaic (PV) generation, storage batteries, electric vehicle (EV) chargers, and controllable loads like traditional air conditioning and heat pumps. These resources can also stress grids due to high feeder voltage and congestion, often during summer heat waves when grid outages can be truly fatal. To solve these problems, distribution systems certainly require more metering and active management. An important and complimentary approach is intelligent control of DERs, where prosumers (consumers with means of production) autonomously manage their own grid withdrawal or injection according to an objective, such as reducing the prosumer's cost of energy or carbon footprint. This unit commitment problem is generally a function of the prosumer's electric load, generation, electric tariff, and dispatchable energy resources. muGrid Analytics, the sponsor of this research, approaches this problem with a techno-economic model of the building energy system and objective function, forecasting consumption and production, and mathematical optimization. The difficulty of the problem increases with, among other things, the complexity of the objective function (driven by the electric tariff) and the uncertainty in knowing future load and renewable generation. Whereas forecasting solar PV and wind generation is a similar problem regardless of scale and is extremely well researched in the literature, effective forecasting building-scale electric load is more of an open question. This research therefore focuses on reviewing existing methods and developing new computational intelligence techniques for load forecasting, especially at the single building or campus scale, considering the forecast must be effective in increasing the performance of muGrid's optimal energy management system (EMS), Redcloud. First, the EMS and optimizer is thoroughly investigated in order to understand the properties of effective load forecasts. A forecast's temporal error distribution is found to have a strong impact on the EMS's ability to reduce the cost of energy, in particular errors clustered around peak load or peak price events can be particularly damaging, but much less so when the errors are a constant offset and less stochastic. Next, an expansive review of the academic literature and industrial state of the art highlights modest academic agreement on the best approaches, complex machine learning or artificial intelligence models don't always substantially outperform traditional methods like exponential smoothing and SARIMA, a substantial benefit to having large data sets, and a need for forecasts to also provide information on the uncertainty. Pre-processing data for better feature extraction is clearly advantageous, including methods like empirical mode decomposition (EMD). Ensemble methods are generally effective because they combine positive attributes of several or many different models. One of the most well-proven machine algorithms is long short-term memory (LSTM) neural networks, which carefully handle relevant information inside of time series. It is so far unclear if large foundational models will consistently outperform their smaller and more custom counterparts, or if the benefit is enough to merit the large storage and computational resource required. A very new and promising ensemble method is AutoTs, winner of the M6 forecasting competition for portfolio management performance, is implemented and tested with moderate results. Therefore, a novel method is developed and tested by this research called ELBA: EMD-LSTM with Bayesian inference and Attention. It is a composite algorithm which takes a well-proven machine learning algorithm, LSTM, pre-processes the inputs using EMD, and adds one of the most important features of modern large language and reasoning models: attention. To increase robustness and provide forecast uncertainty, the method uses a novel implementation of Bayes theorem to seasonal time series, which also is used in ensemble to incorporate other features or forecasts such as seasonal persistence. The complete ELBA method is tested in simulation on an industrial load data set, achieving a skill score (relative improvement over a benchmark method) of 40.7% compared to seasonal persistence. Furthermore the skill score was positive, meaning the method had a lower error than the benchmark, for 96% of forecasted days. The method was also validated online in vivo at the Politecnico di Milano MG2Lab microgrid, achieving agreement with the equivalent simulated data to within €0.43, or 0.15%, of the total cost of energy.
La transizione verso l'energia rinnovabile è in grado di democratizzare la produzione di energia e sbloccare capitali privati adottando le risorse energetiche distribuite (DER). Queste spesso includono la generazione solare fotovoltaica (FV), le batterie di accumulo, i caricatori per veicoli elettrici (EV) e i carichi controllabili come l'aria condizionata tradizionale e le pompe di calore. Queste risorse possono anche stressare le reti a causa dell'alta tensione degli circuiti e della congestione, spesso durante le ondate di calore estive quando le interruzioni di rete possono essere veramente fatali. Per risolvere questi problemi, i sistemi di distribuzione richiedono certamente più contatori e una gestione attiva. Un approccio importante e complementare è il controllo intelligente delle DER, in cui i prosumer (consumatori con fonti di produzione) gestiscono autonomamente il proprio prelievo o immissione in rete secondo un obiettivo, come la riduzione del costo o del CO2 dell'energia. Questo problema di unit commitment è generalmente una funzione del carico elettrico, della generazione, della tariffa elettrica e delle risorse energetiche dispacciabili del prosumer. muGrid Analytics, lo sponsor di questa ricerca, affronta questo problema con un modello tecno-economico del sistema energetico dell'edificio e della funzione obiettivo, prevedendo consumo e produzione, e con l'ottimizzazione matematica. La difficoltà del problema aumenta, tra le altre cose, con la complessità della funzione obiettivo (determinata dalla tariffa elettrica) e con l'incertezza del carico futuro e della generazione rinnovabile. Mentre la previsione della generazione solare fotovoltaica ed eolica è un problema simile indipendentemente dalla scala ed è estremamente ben studiato in letteratura, la previsione efficace del carico elettrico edile è una questione più aperta. Questa ricerca si concentra quindi sulla ricerca dei metodi esistenti e sullo sviluppo di nuove tecniche di intelligenza computazionale per la previsione del carico, specialmente su scala di singolo edificio o campus, considerando che la previsione deve essere efficace nell'aumentare le prestazioni del sistema di gestione energetica ottimale (EMS) di muGrid, Redcloud. Prima, l'EMS e l'ottimizzatore vengono analizzati a fondo per comprendere le dinamiche delle previsioni efficaci. Si è scoperto che la distribuzione temporale dell'errore di una previsione ha un forte impatto sulla capacità dell'EMS di ridurre il costo dell'energia; in particolare, gli errori raggruppati intorno al carico di picco o agli eventi di prezzo di picco possono essere particolarmente dannosi, ma molto meno quando gli errori sono uno scostamento costante e meno stocastici. Successivamente, un'ampia ricerca della letteratura accademica e dello stato dell'arte industriale evidenzia un modesto accordo accademico sui migliori approcci, modelli complessi di machine learning o intelligenza artificiale non sempre superano sostanzialmente i metodi tradizionali come il exponential smoothing e SARIMA, un notevole vantaggio nell'avere grandi set di dati e la necessità che le previsioni forniscano anche informazioni sull'incertezza. La elaborazione dei dati per una migliore estrazione delle caratteristiche è chiaramente vantaggiosa, includendo metodi come la decomposizione modale empirica (EMD). I metodi d'insieme sono generalmente efficaci perché combinano gli attributi positivi di diversi o molti modelli differenti. Uno degli algoritmi di machine learning più collaudati sono le reti neurali a memoria a lungo e breve termine (LSTM), che gestiscono attentamente le informazioni rilevanti all'interno delle serie temporali. Non è ancora chiaro se i grandi modelli fondazionali supereranno costantemente le loro controparti più piccole e su misura, o se il beneficio sia sufficiente a meritare le grandi risorse di memorizzazione e computazionali richieste. Un metodo d'insieme molto nuovo e promettente è AutoTs, vincitore della competizione di previsione M6 per rendimento di gestione del portafoglio, che è stato implementato e testato con risultati moderati. Pertanto, in questa ricerca viene sviluppato e testato un metodo innovativo chiamato ELBA: EMD-LSTM con inferenza Bayesiana e Attenzione. È un algoritmo composito che prende un algoritmo di machine learning ben collaudato, LSTM, pre-elabora gli input utilizzando l'EMD e aggiunge una delle caratteristiche più importanti dei moderni modelli linguistici e di ragionamento di grandi dimensioni: l'attenzione. Per aumentare la robustezza e fornire l'incertezza della previsione, il metodo utilizza un'implementazione innovativa del teorema di Bayes per le serie temporali stagionali, che viene anche utilizzata in un insieme per incorporare altre caratteristiche o previsioni come la persistenza stagionale. Il metodo ELBA completo è testato in simulazione su un set di dati di carico industriale, ottenendo uno skill score (miglioramento relativo rispetto a un metodo di riferimento) del 40,7% rispetto alla persistenza stagionale. Inoltre, lo skill score è stato positivo, il che significa che il metodo ha avuto un errore inferiore rispetto al benchmark, per il 96% dei giorni previsti. Il metodo è stato anche validato online in vivo) presso la microgrid MG2Lab del Politecnico di Milano, ottenendo una concordanza con i dati simulati equivalenti entro 0,43 €, pari allo 0,15%, del costo totale dell'energia.
Computational intelligence for electric load forecasting and microgrid energy management
WOOD, MICHAEL JAMES
2024/2025
Abstract
The renewable energy transition is able to democratize energy production and unlock vast private capital by embracing distributed energy resources (DERs). These often include solar photovoltaic (PV) generation, storage batteries, electric vehicle (EV) chargers, and controllable loads like traditional air conditioning and heat pumps. These resources can also stress grids due to high feeder voltage and congestion, often during summer heat waves when grid outages can be truly fatal. To solve these problems, distribution systems certainly require more metering and active management. An important and complimentary approach is intelligent control of DERs, where prosumers (consumers with means of production) autonomously manage their own grid withdrawal or injection according to an objective, such as reducing the prosumer's cost of energy or carbon footprint. This unit commitment problem is generally a function of the prosumer's electric load, generation, electric tariff, and dispatchable energy resources. muGrid Analytics, the sponsor of this research, approaches this problem with a techno-economic model of the building energy system and objective function, forecasting consumption and production, and mathematical optimization. The difficulty of the problem increases with, among other things, the complexity of the objective function (driven by the electric tariff) and the uncertainty in knowing future load and renewable generation. Whereas forecasting solar PV and wind generation is a similar problem regardless of scale and is extremely well researched in the literature, effective forecasting building-scale electric load is more of an open question. This research therefore focuses on reviewing existing methods and developing new computational intelligence techniques for load forecasting, especially at the single building or campus scale, considering the forecast must be effective in increasing the performance of muGrid's optimal energy management system (EMS), Redcloud. First, the EMS and optimizer is thoroughly investigated in order to understand the properties of effective load forecasts. A forecast's temporal error distribution is found to have a strong impact on the EMS's ability to reduce the cost of energy, in particular errors clustered around peak load or peak price events can be particularly damaging, but much less so when the errors are a constant offset and less stochastic. Next, an expansive review of the academic literature and industrial state of the art highlights modest academic agreement on the best approaches, complex machine learning or artificial intelligence models don't always substantially outperform traditional methods like exponential smoothing and SARIMA, a substantial benefit to having large data sets, and a need for forecasts to also provide information on the uncertainty. Pre-processing data for better feature extraction is clearly advantageous, including methods like empirical mode decomposition (EMD). Ensemble methods are generally effective because they combine positive attributes of several or many different models. One of the most well-proven machine algorithms is long short-term memory (LSTM) neural networks, which carefully handle relevant information inside of time series. It is so far unclear if large foundational models will consistently outperform their smaller and more custom counterparts, or if the benefit is enough to merit the large storage and computational resource required. A very new and promising ensemble method is AutoTs, winner of the M6 forecasting competition for portfolio management performance, is implemented and tested with moderate results. Therefore, a novel method is developed and tested by this research called ELBA: EMD-LSTM with Bayesian inference and Attention. It is a composite algorithm which takes a well-proven machine learning algorithm, LSTM, pre-processes the inputs using EMD, and adds one of the most important features of modern large language and reasoning models: attention. To increase robustness and provide forecast uncertainty, the method uses a novel implementation of Bayes theorem to seasonal time series, which also is used in ensemble to incorporate other features or forecasts such as seasonal persistence. The complete ELBA method is tested in simulation on an industrial load data set, achieving a skill score (relative improvement over a benchmark method) of 40.7% compared to seasonal persistence. Furthermore the skill score was positive, meaning the method had a lower error than the benchmark, for 96% of forecasted days. The method was also validated online in vivo at the Politecnico di Milano MG2Lab microgrid, achieving agreement with the equivalent simulated data to within €0.43, or 0.15%, of the total cost of energy.| File | Dimensione | Formato | |
|---|---|---|---|
|
PhD_Thesis_Wood_v4.pdf
accessibile in internet per tutti
Descrizione: Michael Wood PhD Thesis
Dimensione
6.6 MB
Formato
Adobe PDF
|
6.6 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/10589/244718