The digitalization of the energy industry has made tremendous energy data available. This data is utilized across the entire energy value chain to provide value for customers and energy providers. One area that has gained recent attention in the energy industry is the electricity load forecasting for better scheduling and bidding on the electricity market. However, the electricity data that is used for forecasting is prone to have anomalies, which can affect the accuracy of forecasts. In this thesis we propose two anomaly detection methods to tackle the issue of anomalies in electricity demand data. We propose Long short-term memory (LSTM) and Feed-forward neural network (FFNN) based methods, and compare their anomaly detection performance on two real-world electricity demand datasets. Our results indicate that the LSTM model tends to produce a more robust behavior than the FFNN model on the dataset with regular daily and weekly patterns. However, there was no significant difference between the performance of the two models when the data was noisy and showed no regular patterns. While our results suggest that the LSTM model is effective when a regular pattern in data is present, the results were not found to be statistically significant to claim superiority of LSTM over FFNN. / Digitaliseringen inom energibranschen har tillgängliggjort enorma mängder energidata. Dessa data används över hela värdekedjan för energisystem i syfte att skapa värde för kunder och energileverantörer. Ett område som nyligen uppmärksammats inom energibranschen är att skapa prognoser för elbelastning för bättre schemaläggning och budgivning på elmarknaden. Data som används för sådana prognoser är dock benägna att ha avvikelser, vilket kan påverka prognosernas noggrannhet. I det här examensarbetet föreslår vi två metoder för detektering av avvikelser för att ta itu med frågan om avvikelser i data för elektricitetsbehov. Vi föreslår metoder baserade på Long short-term memory (LSTM) och Feedforward neural network (FFNN) och jämför dess prestanda att upptäcka avvikelser på två verkliga databanker över elbehovsdata. Våra resultat indikerar att LSTM-modellen tenderar att producera ett mer robust beteende än FFNN-modellen på data med upprepande dagliga samt veckovisa mönster. Det fanns dock ingen signifikant skillnad mellan prestanda för de två modellerna när data inte uppvisade regelbunda mönster. Även om våra resultat antyder att LSTM-modellen är effektiv när ett regelbundet datamönster finns närvarande, var resultaten inte statistiskt signifikanta för att påstå överlägsenhet av LSTM jämfört med FFNN.

In this thesis two main studies are conducted to compare the predictive capabilities of feed-forward neural networks (FFNN) and long short-term memory networks (LSTM) in electricity load forecasting. The first study compares univariate networks using past electricity load, as well as multivariate networks using past electricity load and air temperature, in day-ahead load forecasting using varying lookback periods and sparsity of past observations. The second study compares FFNNs and LSTMs of different complexities (i.e. network sizes) when restrictions imposed by limitations of the real world are taken into consideration. No significant differences are found between the predictive performances of the two neural network approaches. However, adding air temperature as extra input to the LSTM is found to significantly decrease its performance. Furthermore, the predictive performance of the FFNN is found to significantly decrease as the network complexity grows, while the predictive performance of the LSTM is found to increase as the network complexity grows. All the findings considered, we do not find that there is enough evidence in favour of the LSTM in electricity load forecasting. / I denna uppsats beskrivs två studier som jämför feed-forward neurala nätverk (FFNN) och long short-term memory neurala nätverk (LSTM) i prognostisering av elkonsumtion. I den första studien undersöks univariata modeller som använder tidigare elkonsumtion, och flervariata modeller som använder tidigare elkonsumtion och temperaturmätningar, för att göra prognoser av elkonsumtion för nästa dag. Hur långt bak i tiden tidigare information hämtas ifrån samt upplösningen av tidigare information varieras. I den andra studien undersöks FFNN- och LSTM-modeller med praktiska begränsningar såsom tillgänglighet av data i åtanke. Även storleken av nätverken varieras. I studierna finnes ingen skillnad mellan FFNN- och LSTM-modellernas förmåga att prognostisera elkonsumtion. Däremot minskar FFNN-modellens förmåga att prognostisera elkonsumtion då storleken av modellen ökar. Å andra sidan ökar LSTM-modellens förmåga då storkelen ökar. Utifrån dessa resultat anser vi inte att det finns tillräckligt med bevis till förmån för LSTM-modeller i prognostisering av elkonsumtion.

No description available.

Short-term load forecasting (STLF) deals with forecasting of hourly system demand with a lead time ranging from one hour to 168 hours. The basic objective of the STLF is to provide for economic, reliable and secure operation of the power system. This dissertation establishes a new approach to artificial neural network (ANN) based STLF. It first decomposes the prediction problem into representation and function approximation problems. The representation problem is solved using phase-space embedding which identifies time delay variables from load time series that are used in forecasting. The concept is inherently different from the methods used so far because it does not use correlated variables for forecasting. Temperature variables are included as well using identified embedding parameters. Function approximation problem is approached using ANN ensemble and active selection of a training set. Training set is selected based on predicted weather parameters for a prediction horizon. Selection is done applying the k-nearest neighbors technique in a temperature-based vector space. A novel approach of pilot set simulation is used to determine the number of hidden units for every forecast period. Ensemble consists of two ANNs which are trained and cross validated on complementary training sets. Final prediction is obtained by a simple average of two trained ANNs. The described technique is used for predicting one week’s load in four selected months in summer peaking and winter peaking US utilities. Mean absolute percent errors (MAPEs) for 24-hour lead time predictions are slightly greater than 2% for all months. For 120-hour lead time (weekday) predictions, MAPEs are around 2.3%. MAPEs for 48- hour lead time (weekend) predictions are around 2.5%. Maximal errors for these cases are around 7%. Predictions for one-hour lead time are slightly higher than 1% for all months, with maximal errors not exceeding 4.99%. Peak load MAPEs are 2.3% for both utilities. Maximal peak-load errors do not exceed 6%. The technique shows very good performance faced with sudden and large changes in weather. For changes in temperature larger than 20° F for two consecutive days, forecasting error is smaller than 3.58%. / Ph. D.

Metamodeling is regarded as a powerful analysis tool to learn the input-output relationship of a system based on a limited amount of data collected when experiments with real systems are costly or impractical. As a popular metamodeling method, Gaussian process regression (GPR), has been successfully applied to analyses of various engineering systems. However, GPR-based metamodeling for time-dependent systems (TDSs) is especially challenging due to three reasons. First, TDSs require an appropriate account for temporal effects, however, standard GPR cannot address temporal effects easily and satisfactorily. Second, TDSs typically require analytics tools with a sufficiently high computational efficiency to support online decision making, but standard GPR may not be adequate for real-time implementation. Lastly, reliable uncertainty quantification is a key to success for operational planning of TDSs in real world, however, research on how to construct adequate error bounds for GPR-based metamodeling is sparse. Inspired by the challenges encountered in GPR-based analyses of two representative stochastic TDSs, i.e., load forecasting in a power system and trajectory prediction for unmanned aerial vehicles (UAVs), this dissertation aims to develop novel modeling, sampling, and statistical analysis techniques for enhancing the computational and statistical efficiencies of GPR-based metamodeling to meet the requirements of practical implementations. Furthermore, an in-depth investigation on building uniform error bounds for stochastic kriging is conducted, which sets up a foundation for developing robust GPR-based metamodeling techniques for analyses of TDSs under the impact of strong heteroscedasticity. / Ph.D. / Metamodeling has been regarded as a powerful analysis tool to learn the input-output relationship of an engineering system with a limited amount of experimental data available. As a popular metamodeling method, Gaussian process regression (GPR) has been widely applied to analyses of various engineering systems whose input-output relationships do not depend on time. However, GPR-based metamodeling for time-dependent systems (TDSs), whose input-output relationships depend on time, is especially challenging due to three reasons. First, standard GPR cannot properly address temporal effects for TDSs. Second, standard GPR is typically not computationally efficient enough for real-time implementations in TDSs. Lastly, research on how to adequately quantify the uncertainty associated with the performance of GPR-based metamodeling is sparse. To fill this knowledge gap, this dissertation aims to develop novel modeling, sampling, and statistical analysis techniques for enhancing standard GPR to meet the requirements of practical implementations for TDSs. Effective solutions are provided to address the challenges encountered in GPR-based analyses of two representative stochastic TDSs, i.e., load forecasting in a power system and trajectory prediction for unmanned aerial vehicles (UAVs). Furthermore, an in-depth investigation on quantifying the uncertainty associated with the performance of stochastic kriging (a variant of standard GPR) is conducted, which sets up a foundation for developing robust GPR-based metamodeling techniques for analyses of more complex TDSs.

Emerging technologies, including distributed energy resources (DERs), internet-of-things and advanced distribution management systems, are revolutionizing the power industry. They provide benefits like higher operation flexibility and lower bulk grid dependency, and are moving the modern power grid towards a decentralized, interconnected and intelligent direction. Consequently, the emphasis of the system operation management has been shifted from the supply-side to the demand-side. It calls for a reconsideration of the business model for future retail market operators. To address this need, this dissertation proposes an innovative retail market solution tailored to market environments penetrated with price-responsive DERs. The work is presented from aspects of theoretical study, test-bed platform development, and experimental analysis, within which two topics relevant to the retail market operation are investigated in depth. The first topic covers the modeling of key retail market participants. With regard to price-insensitive participants, fixed loads are treated as the representative. Deep learning-based day-ahead load forecasting models are developed in this study, utilizing both recurrent and convolutional neural networks, to predict the part of demands that keep fixed regardless of the market price. With regard to price-sensitive participants, battery storages are selected as the representative. An optimization-based battery arbitrage model is developed in this study to represent their price-responsive behaviors in response to a dynamic price. The second topic further investigates how the retail market model and pricing strategy should be designed to incentivize these market participants. Different from existing works, this study innovatively proposes a profit-neutral double-price-signal retail market model. Such a design differentiates elastic prosumers, who actively offer flexibilities to the system operation, from normal inelastic consumers/generators, based on their sensitivities to the market price. Two price signals, namely retail grid service price and retail energy price, are then introduced to separately quantify values of the flexibility, provided by elastic participants, and the electricity commodity, sold/bought to/from inelastic participants. Within the proposed retail market, a non-profit retail market operator (RMO) manages and settles the market through determining the price signals and supplementary subsidy to minimize the overall system cost. In response to the announced retail grid service price, elastic prosumers adjust their day-ahead operating schedules to maximize their payoffs. Given the interdependency between decisions made by the RMO and elastic participants, a retail pricing scheme, formulated based on a bi-level optimization framework, is proposed. Additional efforts are made on merging and linearizing the original non-convex bi-level problem into a single-level mixed-integer linear programming problem to ensure the computational efficiency of the retail pricing tool. Case studies are conducted on a modified IEEE 34-bus test-bed system, simulating both physical operations of the power grid and financial interactions inside the retail market. Experimental results demonstrate promising properties of the proposed retail market solution: First of all, it is able to provide cost-saving benefits to inelastic customers and create revenues for elastic customers at the same time, justifying the rationalities of these participants to join the market. Second of all, the addition of the grid service subsidy not only strengthens the profitability of the elastic customer, but also ensures that the benefit enjoyed per customer will not be compromised by the competition brought up by a growing number of participants. Furthermore, it is able to properly capture impacts from line losses and voltage constraints on the system efficiency and stability, so as to derive practical pricing solutions that respect the system operating rules. Last but not least, it encourages the technology improvement of elastic assets as elastic assets in better conditions are more profitable and could better save the electricity bills for inelastic customers. Above all, the superiority of the proposed retail market solution is proven. It can serve as a promising start for the retail electricity market reconstruction. / Doctor of Philosophy / The electricity market plays a critical role in ensuring the economic and secure operation of the power system. The progress made by distributed energy resources (DERs) has reshaped the modern power industry bringing a larger proportion of price-responsive behaviors to the demand-side. It challenges the traditional wholesale-only electricity market and calls for an addition of retail markets to better utilize distributed and elastic assets. Therefore, this dissertation targets at offering a reliable and computational affordable retail market solution to bridge this knowledge gap. Different from existing works, this study assumes that the retail market is managed by a profit-neutral retail market operator (RMO), who oversees and facilitates the system operation for maximizing the system efficiency rather than making profits. Market participants are categorized into two groups: inelastic participants and elastic participants, based on their sensitivity to the market price. The motivation behind this design is that instead of treating elastic participants as normal customers, it is more reasonable to treat them as grid service providers who offer operational flexibilities that benefit the system efficiency. Correspondingly, a double-signal pricing scheme is proposed, such that the flexibility, provided by elastic participants, and the electricity commodity, generated/consumed by inelastic participants, are separately valued by two distinct prices, namely retail grid service price and retail energy price. A grid service subsidy is also introduced in the pricing system to provide supplementary incentives to elastic customers. These two price signals in addition to the subsidy are determined by the RMO via solving a bi-level optimization problem given the interdependency between the prices and reaction of elastic participants. Experimental results indicate that the proposed retail market model and pricing scheme are beneficial for both types of market participants, practical for the network-constrained real-world implementation, and supportive for the technology improvement of elastic assets.

Published Article / Several forecasting models are available for research in predicting the shape of electric load curves. The development of Artificial Intelligence (AI), especially Artificial Neural Networks (ANN), can be applied to model short term load forecasting. Because of their input-output mapping ability, ANN's are well-suited for load forecasting applications. ANN's have been used extensively as time series predictors; these can include feed-forward networks that make use of a sliding window over the input data sequence. Using a combination of a time series and a neural network prediction method, the past events of the load data can be explored and used to train a neural network to predict the next load point. In this study, an investigation into the use of ANN's for short term load forecasting for Bloemfontein, Free State has been conducted with the MATLAB Neural Network Toolbox where ANN capabilities in load forecasting, with the use of only load history as input values, are demonstrated.

[pt] A previsão de curto prazo, que envolve dados de alta frequência, é essencial para a confiabilidade e eficiência da operação do setor elétrico, fazendo com que a alocação da carga seja feita de forma eficiente, além de indicar possíveis distorções nos próximos períodos (dias, horas, ou frações de hora). A fim de garantir a operação energética, diversas abordagens têm sido empregadas com vistas à previsão de carga de energia a curto prazo. Dentre elas, pode-se citar os modelos híbridos de Séries Temporais, Lógica Fuzzy e Redes Neurais e o Método Holt-Winters com múltiplos ciclos que é a principal ferramenta utilizada atualmente. O HPA (Hierarchical Profiling Approach) é um modelo que decompõe a variabilidade dos dados de séries temporais em três componentes: determinística, estocástica e ruído. A metodologia é capaz de tratar observações únicas, periódicas e aperiódicas, e ao mesmo tempo, serve como uma técnica de pré-branqueamento. Este trabalho tem por objetivo implementar o HPA e aplicá-lo a dados de carga de energia elétrica de 15 em 15 minutos pra um estado da região Sudeste do Brasil. Também serão analisadas as previsões de curto prazo geradas pelo modelo para a série considerada, visto que a habilidade preditiva do HPA ainda é desconhecida para séries brasileiras. As previsões forneceram Coeficiente U de Theil igual a 0,36 e um Erro Percentual Absoluto Médio (MAPE, Mean Absolute Percentage Error) de 5,46%, o qual é bem inferior ao valor fornecido pelo Modelo Ingênuo usado para comparação (15,08%). / [en] Short-term forecast, which involves high frequency data, is essential for a reliable and efficient electricity sector operation, enabling an efficient power load allocation and indicating possible distortions in the coming periods (days, hours, or hour fractions). To ensure the operation efficiency, several approaches have been employed in order to forecast the short-term load. Among them, one can mention the hybrid models of Time Series, Fuzzy Logic and Neural Networks and Holt-Winters Method with multiple cycles, which is the main tool used today. The HPA (Hierarchical Profiling Approach) model decomposes the variability of time series data into three components: deterministic, stochastic and noise. The model is capable of modeling single, periodic and aperiodic observations, and at the same time function as a pre-whitening technique. This work aims to implement the HPA and to apply it in 15 in 15 minutes load data of a Brazil’s southeastern state, since the predictive ability of the HPA is still not known for the Brazilian series. The short-term forecasts estimated for the series considered are analyzed and provided a Theil-U Coefficient equal to 0.36 and a Mean Absolute Percentage Error (MAPE) of 5.46%, which is smaller than the value given by the Naive Model (15.08%).

[pt] Essa tese considera um modelo não linear para se obter previsões de curto prazo de carga de energia elétrica. O modelo combina um modelo de múltiplos regimes auto-regressivo com transição suave com um periódico auto-regressivo criando o modelo de múltiplos regimes periódico com transição suave (STPAR). Um método de construção do modelo é desenvolvido com métodos estatísticos simples e um teste de linearidade contra a hipótese de modelo periódico autoregressivo com transição suave. Outros dois destes foram elaborados para se avaliar o modelo estimado: um teste de Multiplicador de Lagrange (LM) para a hipótese de auto-correlação serial dos resíduos e outro teste LM para a hipótese de não linearidade remanescente. Um experimento de Monte Carlo foi implementado para avaliar a performance dos testes propostos. Estimação por mínimos quadrados não lineares é considerado. Finalmente, dados de carga de energia elétrica do estado de New South Wales na Austrália são apresentados e foram usados como exemplo real. Outros modelos foram utilizados para comparar a performance do modelo. / [en] This thesis considers a non linear approach to obtain short term forecast for electricity load. The model combines a smooth transition autoregressive process with a periodic autoregressive time series model, creating the Smooth Transition Periodic Autoregressive (STPAR) model. A model-building procedure is developed and a linearity test against smooth transition periodic auto-regressive is proposed. Other two tests were created to evaluate the model: a Lagrange multiplier (LM) test for the hypothesis of no error autocorrelation and LM-type test for the hypothesis of no remaining non-linearity. A Monte Carlo experiment was implemented to evaluate the performance of the proposed tests. Estimation by nonlinear least squares is considered. Finally, load data from New South Wales State in Australia`s electricity retail market is presented and will be used as a real example. Other models were used to compare the performance of the proposes model.

Hoje em dia, com a privatização e aumento da competitividade no mercado elétrico, as empresas precisam encontrar formas de melhorar a qualidade do serviço e garantir lucratividade. A previsão de carga de curto prazo é uma atividade indispensável à operação que pode melhorar a segurança e diminuir custos de geração. A fim de realizar a previsão da carga, é necessária a identificação de padrões de comportamento de consumo e da sua relação com variáveis exógenas ao sistema tais como condições climáticas. Originalmente o problema foi resolvido de forma matemática e estatística através de técnicas tais como as séries numéricas, que fornecem bons resultados, mas utilizam processos complexos e de difícil modelamento. O surgimento das técnicas de inteligência artificial forneceu uma nova ferramenta capaz de lidar com a grande massa de dados das cargas e inferir por si mesmo a relação entre as variáveis do sistema. Notadamente, as redes neurais e a lógica fuzzy se destacaram como as técnicas mais adequadas, sendo que já vem sendo estudadas e utilizadas para a previsão de carga a mais de 20 anos. Este trabalho apresenta uma metodologia para a previsão da curva de carga diária de áreas elétricas através do uso de técnicas de inteligência artificial, mais especificamente as redes neurais. Inicialmente são apresentadas as principais técnicas de previsão sendo dado maior detalhamento as redes neurais e a lógica fuzzy. É feita a análise dos dados necessários à previsão e seu tratamento. Em seguida, o processo do uso de redes neurais e lógica fuzzy na previsão é descrito e é apresentado o desenvolvimento e resultados obtidos com o desenvolvimento e implementação de um sistema de previsão com redes neurais na concessionária CTEEP Transmissão Paulista. Como contribuição dessa tese, a transformada Wavelet é analisada como ferramenta para a filtragem e compactação de dados na previsão com redes neurais. / Nowadays, with privatization of utility companies and increase in competition in the energy market, companies must increase their service quality and ensure profits. Short term load forecasting is essential for operation of power systems and can increases security and reduces generation costs. Forecasting the load demands the identification of load patterns and its relations with exogenous variables such as weather. Originally, the problem was solved using mathematics and statistics with techniques such as time series, which produces good results but are complex and have a difficult modeling. With the advent of artificial intelligence techniques, new tools capable of dealing with large amounts of data and learn by themselves system variables relations were available. Artificial neural networks and fuzzy logic came up as the most suitable for load forecasting that have been tested and used for load forecasting for the last 20 years. This work presents a methodology for daily load forecasting of electrical areas using artificial intelligence techniques, specifically neural networks. At first, forecasting techniques are presented with emphasis on neural networks and fuzzy logic. Acquisition and treatment of data are analyzed. The load forecasting using neural networks and fuzzy logic is described and the results of the development and tests of a load forecasting system at CTEEP Transmissão Paulista presented. As contribution of this thesis, Wavelet transform is analyzed as a tool for denoising and data compression for neural network load forecasting.