"The adoption of electricity generation from renewable sources, as well as the push for a speedy electrification of sectors such as transportation and buildings, makes peak electric load management an essential aspect to ensure the electric grid’s reliability and safety. Utilities have established peak load charges that can amount to up to 70% of electricity costs to transfer the financial burden of managing these loads to the consumers. These pricing schemes have created a need for efficient peak electric load management strategies that consumers can implement in order to reduce the financial impact of this type of load. Research has shown that the impact of peak load charges can be reduced by acting on the intelligence provided by peak electric load days (PELDs) forecasts. Unfortunately, published PELDs forecasting methodologies have not addressed the increasing number of facilities adopting behind the meter renewable electricity generation. The presence of this type of intermittent generation adds substantial complexity and other challenges to the PELDs forecasting process. The work reported in this dissertation is organized in terms of its three main contributions to the body of knowledge and to society. First, the development and testing of a first of its kind PELDs forecasting methodology able to accurately predict upcoming PELDs for a consumer regardless of the presence or absence of renewable electricity generation. Experimental results showed that 93% and 90% of potential savings (approximately US$ 142,129.01 and US$ 123,100.74) could be achieved by a consumer with and a consumer without behind the meter solar generation respectively. The second contribution is the development and testing of a novel methodology that allows virtually any type of consumer to determine an efficient electricity demand threshold value before the start of a billing period. This threshold value allows consumers to proactively trigger demand response actions and reduce peak demand charges without receiving any type of signal or information from the utility. Experimental results showed 65% to 82% of total potential demand charge reductions achieved during a year for three different consumers: residential, industrial, and educational with solar generation. These results translate to US$ 149.09, US$ 23,290.00, and US$ 107,610.00 in demand charges savings a year respectively. As a third contribution, we present experimental results that show how the implementation of machine learning based ensemble classification techniques improves the PELDs forecasting methodology’s performance beyond previously published ensemble techniques for three different consumers."--Abstract.

The overarching aim of this open access book is to present self-contained theory and algorithms for investigation and prediction of electric demand peaks. A cross-section of popular demand forecasting algorithms from statistics, machine learning and mathematics is presented, followed by extreme value theory techniques with examples. In order to achieve carbon targets, good forecasts of peaks are essential. For instance, shifting demand or charging battery depends on correct demand predictions in time. Majority of forecasting algorithms historically were focused on average load prediction. In order to model the peaks, methods from extreme value theory are applied. This allows us to study extremes without making any assumption on the central parts of demand distribution and to predict beyond the range of available data. While applied on individual loads, the techniques described in this book can be extended naturally to substations, or to commercial settings. Extreme value theory techniques presented can be also used across other disciplines, for example for predicting heavy rainfalls, wind speed, solar radiation and extreme weather events. The book is intended for students, academics, engineers and professionals that are interested in short term load prediction, energy data analytics, battery control, demand side response and data science in general.

"The addition of solar panels to forecasting energy demand and peak energy demand presents an entirely new challenge to a facility. By having to account for the varying energy generation from the solar panels on any given day based on the weather it becomes increasingly difficult to accurately predict energy demand. With renewable energy sources becoming more prevalent, new methods to track peak energy demand are needed to account for the energy provided by renewable sources. We know from previous research that Artificial Neural Networks (ANN) and Auto Regressive Integrated Moving Average (ARIMA) models are both capable of accurately forecasting building demand and peak electric load days without the presence of solar panels. The goal of this research was to take three different approaches for both the ANN model and the ARIMA model to find the most accurate method for forecasting monthly energy demand and peak load days while considering the varying daily solar energy production. The first approach used was to forecast net demand outright based on relevant historical training data including weather information that would help the models learn how this information affected the overall net demand. The second approach was to forecast the building demand specifically based on the same relevant historical data and then use a random decision tree forest to predict the cluster of day that each day of the month would be in terms of solar production (high, medium with early peak, medium with late peak, low). After the type of day was predicted we would subtract the average solar energy production of the predicted cluster to receive our forecasted net demand for that day. The third approach was similar to the second, but instead of subtracting the average of the cluster we subtracted multiple randomly generated days from that cluster to provide multiple overlapping forecasts. This was specifically used to try and better predict peak load days by testing the hypothesis that if 80% or higher predicted a peak day it would in fact be a peak day. The ANN model outperformed the ARIMA for each approach. Forecasting multiple days was the best of the three approaches. The multiple day ANN forecast had the highest balanced accuracy and sensitivity, the net demand ANN approach was the 2nd most accurate approach and the average solar ANN forecast was the 3rd best approach in terms of balanced accuracy and sensitivity. Based on the outcomes of this study, consumers and institutions such as RIT will be better able to predict peak usage days and use preventative measures to save money by reducing their energy intake on those predicted days. Another benefit will be that energy distribution companies will be able to accurately predict the amount of energy customers with personal solar panels will need in addition to the solar energy they are using. This will allow a greater level of reliability from the providers. Being able to accurately forecast energy demand with the presence of solar energy is going to be critical with the ever-increasing usage of renewable energy."--Abstract.

"Academic institutions spend thousands of dollars every month on their electric power consumption. Some of these institutions follow a demand charges pricing structure; here the amount a customer pays to the utility is decided based on the total energy consumed during the month, with an additional charge based on the highest average power load required by the customer over a moving window of time as decided by the utility. Therefore, it is crucial for these institutions to minimize the time periods where a high amount of electric load is demanded over a short duration of time. In order to reduce the peak loads and have more uniform energy consumption, it is imperative to predict when these peaks occur, so that appropriate mitigation strategies can be developed. The research work presented in this thesis has been conducted for Rochester Institute of Technology (RIT), where the demand charges are decided based on a 15 minute sliding window panned over the entire month. This case study makes use of different statistical and machine learning algorithms to develop a forecasting strategy for predicting the peak electric load days of the month. The proposed strategy was tested for a whole year starting May 2015 to April 2016 during which a total of 57 peak days were observed. The model predicted a total of 74 peak days during this period, 40 of these cases were true positives, hence achieving an accuracy level of 70 percent. The results obtained with the proposed forecasting strategy are promising and demonstrate an annual savings potential worth about $80,000 for a single submeter of RIT."--Abstract.

This comprehensive open access book enables readers to discover the essential techniques for load forecasting in electricity networks, particularly for active distribution networks. From statistical methods to deep learning and probabilistic approaches, the book covers a wide range of techniques and includes real-world applications and a worked examples using actual electricity data (including an example implemented through shared code). Advanced topics for further research are also included, as well as a detailed appendix on where to find data and additional reading. As the smart grid and low carbon economy continue to evolve, the proper development of forecasting methods is vital. This book is a must-read for students, industry professionals, and anyone interested in forecasting for smart control applications, demand-side response, energy markets, and renewable utilization.

Maintaining the balance between generation and consumption is at the heart of electricity grid operation. A disruption to this balance can lead to grid overloads, outages, system damage, rising electricity costs or wasted electricity. For this reason, accurate forecasting of load behavior is crucial. In this work, two classes of ML-based algorithms were used for load forecasting: the Hidden Markov Models (HMMs) and the Deep Neural Networks (DNNs), both of which provide stable and more accurate results than the considered benchmark methods. HMMs could be successfully used as a stand-alone predictor with a training based on Maximum Likelihood Estimation (MLE) in combination with a clustering of the training data and an optimized Viterbi algorithm, which are the main differences to other HMM-related load forecasting approaches in the literature. Adaptive online training was developed for DNNs to minimize training times and create forecasting models that can be deployed faster and updated as often as necessary to account for the increasing dynamics in power grids related to the growing share of installed renewables. In addition, the flexible and powerful encoder-decoder architecture was used, which helped to minimize the forecast error compared to simpler DNN architectures such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs) and others.