In transforming the current electricity network, in a so called smart grid, demand forecasting is relevant to processes such as demand management, demand response, distributed generation, among others. For consumers, the replacement of electromechanical meters by electronic meters, enables real-time access to measurement data, providing this data for demand forecasting. The present work focuses on consumers with different profiles, commercial, industrial and institutional, connected to the the distribution network in medium-voltage and loads ranging between a few tens of kilowatts and two megawatts. For these consumers, very short-term demand forecasting (up to 2 hours) will be an important tool for decision making in a dynamic environment, with time-variable energy prices, demand-side management and eventually own generation. With the application of demand forecasting methods to various consumers with different profiles, it is shown that the forecasting methods with better accuracy (lower average error) are variable from consumer to consumer. For one consumer individually, the method with better accuracy is also variable, depending on the hour of the day. Combination of several demand forecasting methods results in similar or better performance compared to using only a single method. A method of selective combination is proposed, in order to eliminate the risk of choosing a unique method, which results are unpredictable. The results of the application of the proposed combination method, on several consumers with different characteristics, demonstrate that selective combination improves the quality of the forecast. / No processo de transformação das redes de energia elétrica atuais, em redes elétricas inteligentes (smart grid), a previsão de demanda é relevante para processos como o gerenciamento da demanda, resposta a demanda, geração distribuída, entre outros. Para os consumidores, a substituição de medidores eletromecânicos, por medidores eletrônicos, possibilita o acesso em tempo real aos dados da medição, disponibilizando estes dados para a previsão de demanda. O presente trabalho focaliza em consumidores de diferentes perfis, comerciais, industriais e institucionais, ligados à rede de distribuição em média tensão, com demandas situadas em uma faixa de algumas dezenas de quilowatts até dois megawatts. Para estes consumidores, uma previsão de demanda para curtíssimo prazo (até 2 horas) será uma importante ferramenta na tomada de decisão em um ambiente dinâmico, com tarifas variáveis no tempo, gerenciamento pelo lado da demanda e eventual geração própria. A partir da aplicação de métodos de previsão de demanda em consumidores de variados perfis de carga, é demonstrado que os métodos de previsão com melhor precisão (menor erro médio) variam de consumidor para consumidor. Para um mesmo consumidor o método de melhor precisão também é variável, sendo dependente da hora do dia. Uma combinação de diversos métodos de previsão de demanda resulta em performance similar ou superior se comparado ao uso de apenas um método individual. Propõe-se um método de combinação seletiva, com o objetivo de eliminar o risco da escolha de um único método, cujos resultados são imprevisíveis. Os resultados da aplicação do método de combinação proposto, em diversos consumidores de características diferentes, demonstram que a combinação seletiva representa uma melhora na qualidade da previsão.

Orientadores: Takaaki Ohishi, Ricardo Menezes Salgado / Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Engenharia Elétrica e de Computação / Made available in DSpace on 2018-08-24T15:46:34Z (GMT). No. of bitstreams: 1 Barreto_MateusNeves_M.pdf: 6008302 bytes, checksum: ae210360a5363404761ca9b3566732ab (MD5) Previous issue date: 2014 / Resumo: O Sistema Elétrico Brasileiro abastece cerca de 97% da demanda de energia nacional. Frente ao extenso território brasileiro, necessita-se de um sistema de transmissão de larga escala, devido as grandes distâncias entre as gerações, das hidroelétricas, e a principal concentração da demanda, no Sudeste brasileiro. Para garantir segurança e economia da operação do Sistema Elétrico Brasileiro são realizadas análises da operação do sistema de geração e transmissão frente às condições de cargas críticas. A ideia é preparar o sistema para suportar as condições mais severas de carga. A curva de carga crítica é calculada para cada mês com discretização horária (ou menor). A mesma é composta pela carga mínima observada num dado mês no período da primeira a oitava hora, e pela carga máxima observada no mês para as horas restantes. Utilizando históricos de demanda pertencentes aos agentes do Setor Elétrico Brasil, foi possível criar um histórico de cinco anos, 60 meses, de curvas de carga crítica. Esses dados foram disponibilizados pelo Operador Nacional do Sistema Elétrico Brasileiro ¿ ONS, em conjunto com o desenvolvimento de um projeto de pesquisa, através de um sistema de suporte a decisão nomeado SysPrev. Nesta dissertação são propostos três modelos para realizar a previsão da curva de carga crítica. Dois modelos utilizam Redes Neurais Artificiais e um modelo utiliza Suavização Exponencial de Holt-Winters (HW). Os resultados obtidos por todos os modelos foram satisfatórios. O modelo de Suavização Exponencial se destacou perante os outros dois modelos atingindo erros médios absolutos próximos a 3%. Esses resultados justificam-se devido às séries históricas de curvas de carga crítica possuírem características de tendência e sazonalidade e o modelo de HW ser projetado especificamente para séries temporais com estas características / Abstract: The Brazilian Power System supplies around 97 % of national energy demand. By reason of the broad Brazilian territory, it requires a transmission system of large scale, due to the large distances between the generations of hydropower and the main concentration of demand that stay in southeastern of Brazil. To ensure security and economy of operation of the Brazilian Electric System are performed analyzes the operation of generation and transmission system due to the conditions of critical loads. The idea is to prepare the system to resist the harshest load conditions. The curve of critical load is calculated for each month with hourly discretization (or less). It's made with the minimum load observed in a given month between the first to eighth hour, and to maximum load observed in the month for the rest of hours. Using the demand agents¿ history belonging to the Brazilian Power System, it was possible to create a record of five years, 60 months, of curves of critical load. These datas were available by the National Operator of the Brazilian Power System as part of the development of a research project, made available by a decision support system named SysPrev. This dissertation proposed three models to perform the forecasting of the critical load curve. Two models using Artificial Neural Networks and one model using Exponential Smoothing Holt-Winters (HW). The results obtained by all the models were satisfactory. The exponential smoothing model stood out against the other two models, this having absolute average errors near 3%. These results are justified due to the historical series of curves of critical load has characteristics of trend and seasonality and the HW model is specifically designed for time series with such characteristics / Mestrado / Engenharia de Computação / Mestre em Engenharia Elétrica

[pt] Neste trabalho apresenta-se o estudo de Previsões de demanda de Energia Elétrica utilizando séries temporais, particularmente a teoria devido a Box & Jenkins. Estuda-se um modelo já existente em uma das cidades proporcionando a hipótese de se estender a validade deste modelo, para cidade de mesmas características onde houver falta de dados. Os dados numéricos utilizados neste estudo são relativos à Centrais Elétricas Fluminense, (CELF). / [en] This paper presents a study of load previsions using chronological series, especially the theory of Box and Jenkins. One model is determined for a city and next a trial is made to extend the vality of this model to other cities with the same characteristics when there is a lack of data. The numerical data use in the work are those of CENTRAIS ELETRICAS FLUMINENSE (CELF).

Context. Heat load forecasting is an important part of district heating optimization. In particular, energy companies aim at minimizing peak boiler usage, optimizing combined heat and power generation and planning base production. To achieve resource efficiency, the energy companies need to estimate how much energy is required to satisfy the market demand. Objectives. We suggest an online machine learning algorithm for heat load forecasting. Online algorithms are increasingly used due to their computational efficiency and their ability to handle changes of the predictive target variable over time. We extend the implementation of online bagging to make it compatible to regression problems and we use the Fast Incremental Model Trees with Drift Detection (FIMT-DD) as the base model. Finally, we implement and incorporate to the algorithm a mechanism that handles missing values, measurement errors and outliers. Methods. To conduct our experiments, we use two machine learning software applications, namely Waikato Environment for Knowledge Analysis (WEKA) and Massive Online Analysis (MOA). The predictive ability of the suggested algorithm is evaluated on operational data from a part of the Karlshamn District Heating network. We investigate two approaches for aggregating the data from the nodes of the network. The algorithm is evaluated on 100 runs using the repeated measures experimental design. A paired T-test is run to test the hypothesis that the the choice of approach does not have a significant effect on the predictive error of the algorithm. Results. The presented algorithm forecasts the heat load with a mean absolute percentage error of 4.77\%. This means that there is a sufficiently accurate estimation of the actual values of the heat load, which can enable heat suppliers to plan and manage more effectively the heat production. Conclusions. Experimental results show that the presented algorithm can be a viable alternative to state-of-the-art algorithms that are used for heat load forecasting. In addition to its predictive ability, it is memory-efficient and can process data in real time. Robust heat load forecasting is an important part of increased system efficiency within district heating, and the presented algorithm provides a concrete foundation for operational usage of online machine learning algorithms within the domain.

[pt] A previsão de perfis de carga elétrica (i.e., das séries de cargas a cada hora de um dia) tem sido freqüentemente tentada por meio de modelos baseados em redes neurais. Os resultados conseguidos por estes modelos, contudo, ainda não são considerados inteiramente convincentes. Há duas razões para ceticismo: em primeiro lugar, os modelos sugeridos geralmente se baseiam em redes que parecem ser complexas demais em relação aos dados que pretendem modelar (isto é, estes modelos parecem estar superparametrizados); em segundo lugar, estes modelos geralmente não são bem validados, pois os artigos que os propõem não comparam o desempenho das redes ao de modelos de referência. Nesta tese, examinamos estes dois pontos por meio de revisões críticas da literatura e de simulações, a fim de verificar se é realmente viável a aplicação de redes neurais à previsão de perfis de carga. Nas simulações, construímos modelos bastante complexos de redes e verificamos empiricamente sua validade, pela comparação de seu desempenho preditivo fora da amostra de treino ao desempenho de vários outros modelos de previsão. Os resultados mostram que as redes, mesmo quando muito complexas, conseguem previsões de perfis mais acuradas do que os modelos tradicionais, o que sugere que elas poderão trazer uma grande contribuição para a solução do problema de previsão de cargas. / [en] Load profile forecasting (i.e., the forecasting of series of hourly loads) has been often attempted by means of models based on neural networks. However, the papers that propose such models are not considered entirely convincing, for at least two reasons. First, the models they propose are usually based on neural networks that seem to be too large in relation to the sample they intend to model (that is, the networks seem to overfit their data). Secondly, most of the models are not properly validated, since the papers do not compare their performances to that of any standard forecasting method. In this thesis, we examine these two points, by means of literature reviews and of simulations, in order to investigate the feasibility of the application of neural networks to the problem of profile forecasting. We build some very complex models based on neural networks, and validate them empirically by comparing their predictive performance out-of-sample, over actual data, to the performance of several other forecasting methods. The results show that neural networks, even when very complex, are able to forecast profiles more accurately than the traditional models, which suggests that they may yet bring large contributions to the solution of the load forecasting problem. / [es] La previsión de los perfiles de carga elétrica (i.e., series de cargas medidas a cada hora de un día) ha sido abordada con frecuencia a través de modelos basados en redes neurales. Los resultados obtenidos por estos modelos, todavía no son considerados enteramente convincentes. Existen dos razones para este escepticismo: en primer lugar, los modelos sugeridos generalmente se basan en redes que parecen ser demasiado complejas en relación a los datos que pretenden modelar (quiere decir, estos modelos parecen estar superparametrizados); en segundo lugar, estos modelos generalmente no son bien evaluados, pués los artículos que los proponen no comparan el desempeño de las redes al de los modelos de referencia. En esta tesis, examinamos estos dos puntos por medio de revisiones críticas de la literatura y de simulaciones, con el objetivo de verificar si es realmente viable la aplicación de redes neurales a la previsión de perfiles de carga. En las simulaciones, construímos modelos de redes bastante complejos y verificamos empíricamente su validez, comparando su desempeño predictivo fuera de la muestra de entrenamiento con el desempeño de varios otros modelos de previsión. Los resultados muestran que las redes, incluso cuando muy complejas, consiguen previsiones de perfiles más precisas que los modelos tradicionales, lo que sugiere que ellas poderián traer una gran contribución para la solución del problema de previsión de cargas.

D.Phil. (Electrical and Electronic Engineering) / Load forecasting is a necessary and an important task for both the electrical consumer and electrical supplier. Whilst many studies emphasize the importance of determining the future demand, few papers address both the forecasting algorithm and computational resources needed to offer a turnkey solution to address the load forecasting problem. The major contribution that, this paper identified is a turnkey load forecasting algorithm. A turnkey forecasting solution is defined by a comprehensive solution that incorporates both the algorithm and processing elements needed to execute the algorithm in the most effective and efficient manner. An electrical consumer, namely the operator of a rapid railway system was faced with a problem of having to forecast the notified network demand and energy consumption. The forecast period was expected to be between a very short term window for maintenance reasons and long term for the requirements warranted by the electrical supplier. The problem was addressed by firstly reviewing the most common forms of load forecasting for which there are two types. These are statistically based methods and methods based upon artificial intelligence. The basic principle of a statistical approach is to approximate or define a curve that best defines the relationship between the load and its parameters. Regression and similar day approach methods use the defined correlation of past values in order to forecast the future behaviour. In other words the future load forecast is forecasted by observing the behaviour of the factors that influenced the load behaviour in the past. The underlying factors that influence the final load may be identified by means of a top down drill down approach. In this way both the load factors and influential variables may be identified. This paper makes use of relevance trees to create a structure of load and influential variables. For a regression forecasting model, the behaviour of the load is modelled according to weather and non-weather variables. The load may be stochastic or deterministic, linear or nonlinear. One of the biggest problems with statistical models is the lack of generality. One model may yield more acceptable results over another model simply because of the sensitivity of the model to one load element that defines the model significantly. Regression type forecast models are an example of this where the elements that define the load are broadly divided into weather and non-weather elements. It is important that the correlation curve reflects the true correlation between the load and its elements. The recursive properties of a statistical based techniques (Kalman filter) allows that the relationship be refined. For methods such as neural networks, the relationship between the load elements that define the future load behaviour is learnt by presenting a series of patterns and then a forecast model is derived. Rigorous mathematical equations are replaced with an artificial neural network where the load curve is learnt. Unlike a statistical based approach (ARMA models), the load does not first need to be defined as a stochastic or deterministic series. In terms of a stochastic approach (non stationery process), the load first would have to be brought to a stationery process. For artificial neural networks, such processes are eliminated and the future forecast is derived faster in terms of a turnkey approach (tested solution). Artificial Neural Networks (ANN) has gained momentum since the eighties. Specifically in the area of forecasting, neural networks have become a common application. In this thesis, data from a railway operator was used to train the neural network and then future data is forecasted. Two embedded processing elements were then evaluated in terms of speed, memory and ability to execute complex mathematical functions (libraries). These were namely a Complex Programmable Logic Device (CPLD) and microcontroller (MCU). The ANN forecasting algorithm was programmed on both a MCU and PLD and compared by means of timing models and hardware platform testing. The most ideal turnkey solution was found to be the ANN algorithm residing on a PLD. The accuracy and speed results surpassed that of a MCU.

The dynamic response of the demand is defined as the time-domain real and reactive power response to a voltage disturbance, and it represents the dynamic load characteristics. This thesis develops a methodology for probabilistic estimation and prediction of dynamic responses of the demand at bulk supply points. The main outcome of the research is being able to predict the contribution of different categories of loads to the total demand mix and their controllability without conducting detailed customer surveys or collecting smart meter data, and to predict the dynamic response of the demand without performing field tests. The prediction of the contributions of different load categories and their controllability and load characteristics in the near future (e.g., day ahead) plays an important role in system analysis and planning, especially in the short-term dispatch and control. However, the research related to this topic is missing in the publically available literature, and an approach needs to be developed to enable the prediction of the participation of different loads in total load mix, their controllability and the dynamic response of the demand. This research contributes to a number of areas, such as load forecasting, load disaggregation and load modelling. First, two load forecasting methodologies which have not been compared before are compared; and based on the results of comparison and considering the actual requirements in this research, a methodology is selected and used to predict both the real and reactive power. Second, a unique methodology for load disaggregation is developed. This methodology enables the estimation of the contributions of different load categories to the total demand mix and their controllability based on RMS measured voltage and real and reactive power. The confidence level of the estimation is also assessed. The methodology for disaggregation is integrated with the load forecasting tool to enable prediction of load compositions and dynamic responses of the demand. The prediction is validated with data collected from real UK power network. Finally, based on the prediction, an example of load shifting is used to demonstrate that different dynamic responses can be obtained based on the availability and redistribution of controllable devices and that load shifting decisions, i.e., demand side management actions, should be made based not only on the amount of demand to be shifted, but also on predicted responses before and after load shifting.

D.Phil / The prediction of the expected transmission network loads as required for transmission network power flow studies, has become very important and much more complex than ten to twenty years ago. Therefore a single forecast is no longer the answer to the problem. The modelling of different long-term electrical forecasts makes it possible to compare a number of different forecasts. The modelling provides the further option that each expected load can be entered as a range and then the developed balancing algorithm checks for consensus (feasibility). If feasibility exists, then the different forecasts are reconciled (a feasible solution is determined). Factors such as international and national market trends, economical cycles, different weather patterns, climate cycles and demographic changes are studied. The factors that have significant impact on the transmission electrical loads are integrated in ten different forecasts. It thus gives more insight into the electrical industry and makes the forecast results more informative and therefore reduces the uncertainty in the future expected loads.

No description available.

<div>As on 2017, US Energy Information Administration (US EIA) claims that 50 % of the total US energy consumption are contributed by Commercial and Industrial (C&I) end-users.</div><div>Most of the energy consumption by these users is in the form of the electric power. Electric utilities, who usually supply the electric power, tend to care about the power consumption profiles of these users mainly because of the scale of consumption and their significant contribution</div><div>towards the system peak. Predicting and managing the peaks of C&I users is crucial both for the users themselves and for utility companies.</div><div>In this research, we aim to understand and predict the daily peaks of individual C&I users. To empirically understand the statistical characteristics of the peaks, we perform an extensive exploratory data analysis using a real power consumption time series dataset. To accurately predict the peaks, we investigate indirect and direct learning approaches. In the indirect approach, daily peaks are identified after forecasting the entire time series for the day whereas in the direct approach, the daily peaks are directly predicted based on the historical data available for different users during different days of the week. The machine learning models used in this research are based on Simple Linear Regression (SLR), Multiple Linear Regression (MLR), and Artificial Neural Networks (ANN).</div>