As Next Generation (NextG) networks become more complex, it has become increasingly necessary to utilize more advanced algorithms to enhance the robustness, autonomy, and reliability of existing wireless infrastructure. One such algorithm is network traffic prediction, playing a crucial role in the efficient operation of real-time and near-real-time network management. The contributions of this thesis are twofold. The first introduces a novel cluster-train-predict framework that leverages domain knowledge to identify unique timeseries sub-behaviors within aggregates of network data. This method produces distributions that are more robust towards changes in the spatio-temporal environment. The ensemble of time-series prediction models trained on these distributions posses a greater affinity towards accurate network prediction, selectively employing learned behaviors to handle sources of time-series data without any prior knowledge of it. This approach tends to improve the ability to accurately forecast network traffic volumes. Secondly, this thesis explains the development and implementation of a modular data pipeline to support the cluster-train-predict framework under a variety of conditions. This setup promotes repeatable and comparable results, facilitating rapid iteration and experimentation on current and future research. The results of this thesis surpass traditional approaches in [1] by up to 60%. Furthermore, the effectiveness of this framework is also validated using two additional time-series datasets [2] and [3], demonstrating the ability of this approach to generalize towards other time-series data and machine learning applications in uncertain environments. / Master of Science / As Next Generation (NextG) networks become more complex, it has become increasingly necessary to utilize more advanced algorithms to enhance the robustness, autonomy, and reliability of in-use wireless infrastructure where network traffic prediction plays a crucial role in the efficient operation of real-time and near real-time network management. The contributions of this thesis are twofold. The first explores a novel cluster-train-predict framework that uses an unsupervised learning approach, specifically time-series K-means clustering, to group similar time-series data. In doing so, this approach identifies unique time-series behaviors within network provider data. Since this approach aims to reduce the variance within each aggregate, models can specialize towards particular network behaviors, becoming better suited for a wider variety of network trends during prediction. Because this framework assigns data to each cluster based on these groupings, the framework can adapt towards changes in network infrastructure or underlying shifts in its environment to forecast with a greater degree of certainty and explainability. This framework can even generalize towards out-of-distribution cases where it has no prior knowledge of a source of time-series data outperforming [1] by up to 60%. This approach tends to improve the ability to accurately forecast network traffic volumes. Secondly, this thesis explains the development and implementation of a modular data pipeline to support the cluster-train-predict framework under a variety of conditions with repeatable and comparable results, facilitating rapid iteration and experimentation on current and future research. The results of the framework are also corroborated on two, additional time-series datasets [2] and [3], demonstrating the ability of this approach to generalize towards time-series data, where this framework can also be applied to other machine learning applications in uncertain environments.

The forecasting of time series data is an integral component for management, planning, and decision making in many domains. The prediction of electricity demand and supply in the energy domain or sales figures in market research are just two of the many application scenarios that require thorough predictions. Many of these domains have in common that they are influenced by the Big Data trend which also affects the time series forecasting. Data sets consist of thousands of temporal fine grained time series and have to be predicted in reasonable time. The time series may suffer from noisy behavior and missing values which makes modeling these time series especially hard, nonetheless accurate predictions are required. Furthermore, data sets from different domains exhibit various characteristics. Therefore, forecast techniques have to be flexible and adaptable to these characteristics. Long-established forecast techniques like ARIMA and Exponential Smoothing do not fulfill these new requirements. Most of the traditional models only represent one individual time series. This makes the prediction of thousands of time series very time consuming, as an equally large number of models has to be created. Furthermore, these models do not incorporate additional data sources and are, therefore, not capable of compensating missing measurements or noisy behavior of individual time series. In this thesis, we introduce CSAR (Cross-Sectional AutoRegression Model), a new forecast technique which is designed to address the new requirements on forecasting large-scale time series data. It is based on the novel concept of cross-sectional forecasting that assumes that time series from the same domain follow a similar behavior and represents many time series with one common model. CSAR combines this new approach with the modeling concept of ARIMA to make the model adaptable to the various properties of data sets from different domains. Furthermore, we introduce auto.CSAR, that helps to configure the model and to choose the right model components for a specific data set and forecast task. With CSAR, we present a new forecast technique that is suited for the prediction of large-scale time series data. By representing many time series with one model, large data sets can be predicted in short time. Furthermore, using data from many time series in one model helps to compensate missing values and noisy behavior of individual series. The evaluation on three real world data sets shows that CSAR outperforms long-established forecast techniques in accuracy and execution time. Finally, with auto.CSAR, we create a way to apply CSAR to new data sets without requiring the user to have extensive knowledge about our new forecast technique and its configuration.

Foreign exchange (FX) rate forecasting has been a challenging area of study in the past. Various linear and nonlinear methods have been used to forecast FX rates. As the currency data are nonlinear and highly correlated, forecasting through nonlinear dynamical systems is becoming more relevant. The nearest neighbor (NN) algorithm is one of the most commonly used nonlinear pattern recognition and forecasting methods that outperforms the available linear forecasting methods for the high frequency foreign exchange data. The basic idea behind the NN is to capture the local behavior of the data by selecting the instances having similar dynamic behavior. The most relevant k number of histories to the present dynamical structure are the only past values used to predict the future. Due to this reason, NN algorithm is also known as the k-nearest neighbor algorithm (k-NN). Here k represents the number of chosen neighbors. In the k-nearest neighbor forecasting procedure, similar instances are captured through a distance function. Since the forecasts completely depend on the chosen nearest neighbors, the distance plays a key role in the k-NN algorithm. By choosing an appropriate distance, we can improve the performance of the algorithm significantly. The most commonly used distance for k-NN forecasting in the past was the Euclidean distance. Due to possible correlation among vectors at different time frames, distances based on deterministic vectors, such as Euclidean, are not very appropriate when applying for foreign exchange data. Since Mahalanobis distance captures the correlations, we suggest using this distance in the selection of neighbors. In the present study, we used five different foreign currencies, which are among the most traded currencies, to compare the performances of the k-NN algorithm with traditional Euclidean and Absolute distances to performances with the proposed Mahalanobis distance. The performances were compared in two ways: (i) forecast accuracy and (ii) transforming their forecasts in to a more effective technical trading rule. The results were obtained with real FX trading data, and the results showed that the method introduced in this work outperforms the other popular methods. Furthermore, we conducted a thorough investigation of optimal parameter choice with different distance measures. We adopted the concept of distance based weighting to the NN and compared the performances with traditional unweighted NN algorithm based forecasting. Time series forecasting methods, such as Auto regressive integrated moving average process (ARIMA), are widely used in many ares of time series as a forecasting technique. We compared the performances of proposed Mahalanobis distance based k-NN forecasting procedure with the traditional general ARIM- based forecasting algorithm. In this case the forecasts were also transformed into a technical trading strategy to create buy and sell signals. The two methods were evaluated for their forecasting accuracy and trading performances. Multi-step ahead forecasting is an important aspect of time series forecasting. Even though many researchers claim that the k-Nearest Neighbor forecasting procedure outperforms the linear forecasting methods for financial time series data, and the available work in the literature supports this claim with one step ahead forecasting. One of our goals in this work was to improve FX trading with multi-step ahead forecasting. A popular multi-step ahead forecasting strategy was adopted in our work to obtain more than one day ahead forecasts. We performed a comparative study on the performance of single step ahead trading strategy and multi-step ahead trading strategy by using five foreign currency data with Mahalanobis distance based k-nearest neighbor algorithm.

[pt] A previsão de séries temporais é um assunto de grande importância para diversas áreas, podendo servir como base para planejamento e controle, entre outros. As formas mais comuns de previsão são as pontuais. É arriscado, no entanto, planejadores tomarem decisões unicamente com base em previsões pontuais, pois séries reais são compostas por uma parte aleatória que não pode ser definida por modelagem matemática. Um modo de contornar este problema é realizando previsões intervalares. Estas fornecem informações sobre as incertezas das previsões pontuais, o que auxilia o planejador em suas decisões. Modelos de lógica fuzzy têm sido investigados na literatura de previsão devido a sua capacidade de modelar incertezas. Apesar disso, sistemas de lógica fuzzy Mamdani (MFLS) foram pouco investigados no tema, comparando-se a outros tipos de modelagens fuzzy. Ademais, entende-se que a literatura de previsão intervalar com modelos fuzzy é limitada. Neste contexto, este trabalho propõe um método para construção de previsões intervalares a partir das previsões pontuais do modelo MFLS de tipo-1 (T1 MFLS). O método proposto para construção de previsões intervalares do MFLS é baseado na reamostragem de erros in-sample. O modelo T1 MFLS é construído com uma heurística (para partição do universo de discurso das variáveis do modelo) e com a seleção da entrada do modelo. Previsões pontuais e intervalares são produzidas para séries horárias de carga de energia elétrica. A literatura de modelos fuzzy de previsão é revisada. / [en] Time series forecasting is an important subject for many areas; it can serve as basis for planning and control, among others. The most common type of forecast is the point forecast. It is, nevertheless, risky to make decisions based on point forecasts, considering that real time series are composed by a random part that cannot be exactly defined by mathematical modeling. One way to by-pass this problem is by producing interval forecasts. These provide information about point forecasts reliability, what helps the planner make his decisions. Fuzzy logic models have been investigated in the forecasting literature due to their ability to model uncertainties. In spite of this, Mamdani fuzzy logic systems (MFLS) have been less investigated in this subject than other types of fuzzy modeling approaches. Furthermore, it is understood that the literature of interval forecasting with fuzzy models is very limited. In this context, this work proposes a method for creating interval prediction from point forecasts of a type-1 MFLS (T1 MFLS). The proposed method for interval forecast construction is based on the resampling of in-sample errors. The T1 MFLS model is constructed with a heuristic (that makes the partition of the universe of discourse of the model s variables) and with selection of the model s inputs. Point and interval forecasts are produced for hourly electricity load series. The literature of fuzzy models applied in forecasting is reviewed.

<p>This thesis describes one promising approach where a problem of time<br />series analysis and prediction was solved by using Case Based Reasoning<br />(CBR) technology. Foundations and main concepts of this technology are<br />described in detail. Furthermore, a detailed study of different approaches in<br />time series analysis is given. System CuBaGe (Curve Base Generator) - A<br />robust and general architecture for curve representation and indexing time<br />series databases, based on Case based reasoning technology, was<br />developed. Also, a corresponding similarity measure was modelled for a<br />given kind of curve representation. The presented architecture may be<br />employed equally well not only in conventional time series (where all<br />values are known), but also in some non-standard time series (sparse,<br />vague, non-equidistant). Dealing with the non-standard time series is the<br />highest advantage of the presented architecture.</p> / <p>U ovoj doktorskoj disertaciji prikazan je interesantan i perspektivan pristup<br />re&scaron;avanja problema analize i predviđanja vremenskih serija kori&scaron;ćenjem<br />Case Based Reasoning (CBR) tehnologije. Detaljno su opisane osnove i<br />glavni koncepti ove tehnologije. Takođe, data je komparativna analiza<br />različitih pristupa u analizi vremenskih serija sa posebnim osvrtom na<br />predviđanje. Kao najveći doprinos ove disertacije, prikazan je sistem<br />CuBaGe (Curve Base Generator) u kome je realizovan originalni način<br />reprezentacije vremenskih serija zajedno sa, takođe originalnom,<br />odgovarajućom merom sličnosti. Robusnost i generalnost sistema<br />ilustrovana je realnom primenom u domenu finansijskog predviđanja, gde<br />je pokazano da sistem jednako dobro funkcioni&scaron;e sa standardnim, ali i sa<br />nekim nestandardnim vremenskim serijama (neodređenim, retkim i<br />neekvidistantnim).</p>

Autoscalers handle the scaling of instances in a system automatically based on specified thresholds such as CPU utilization. Reactive autoscalers do not take the delay of initiating a new instance into account, which may lead to overutilization. By applying machine learning methodology to predict future loads and the desired number of instances, it is possible to preemptively initiate scaling such that new instances are available before demand occurs. Leveraging efficient scaling policies keeps the costs and energy consumption low while ensuring the availability of the system. In this thesis, the predictive capability of different multilayer perceptron configurations is investigated to elicit a suitable model for a telecom support system. The results indicate that it is possible to accurately predict future load using a multilayer perceptron regressor model. However, the possibility of reproducing the results in a live environment is questioned as the dataset used is derived from a simulation.

Decentralization of the electric grid can increase resiliency (during natural disasters) and can reduce T&D energy losses and emissions. Microgrids and DERs can enable this to happen. It is important to optimally control microgrids and DERs to extract the greatest economic, environmental and resiliency benefits. This is enabled by robust forecasting to optimally control loads and energy sources. An integral part of microgrid control is power side and load side demand forecasting. In this thesis, we look at the ability of a powerful neural network algorithm to forecast the load side demand for a microgrid using the UMass campus as the test bed. UMass has its own power plant producing 16 MW of power. In addition to this, Solar panels totaling 5.5MW and lithium ion battery bank of 1.32 MW/4 MWh are also available. An LSTM recurrent neural network is used for demand forecasting. In addition to a fully trained LSTM network, multi linear regression model, ARIMA and ANN model are also tested to compare the performance. In addition to the Short Term Load Forecasting, the peak prediction accuracy of the model was also tested to run a battery discharge algorithm to shave peak demand for the microgrid. This will result in demand cost savings for the facility. Finally, the fully trained neural network was deployed on a raspberry pi computer.

[pt] Realizar previsões de preços no mercado de ações é uma tarefa difícil devido ao fato de o mercado financeiro ser um ambiente altamente dinâmico, complexo e caótico. Para algumas teorias financeiras, usar as informações disponíveis para tentar prever o preço de uma ação a curto prazo é um esforço em vão já que ele sofre a influência de diversos fatores externos e, em decorrência, sua variação assemelha-se à de um passeio aleatório. Estudos recentes, como (37) e (51), abordam o problema com modelos de predição específicos para o comportamento do preço de uma ação isolada. Neste trabalho, apresenta-se uma proposta para prever variações de preço tendo como base conjuntos de ações consideradas similares. O objetivo é criar um modelo capaz de prever se o preço de diferentes ações tendem a subir ou não a curto prazo, considerando informações de ações pertencentes a conjuntos similares com base em duas fontes de informações: os dados históricos das ações e as notícias do Google Trends. No estudo proposto, primeiramente é aplicado um método para identificar conjuntos de ações similares para então criar um modelo de predição baseado em redes neurais LSTM (long shortterm memory) para esses conjuntos. Mais especificamente, foram conduzidos dois experimentos: (1) aplicação do algoritmo K-Means para a identificação dos conjuntos de ações similares, seguida da utilização de uma rede neural LSTM para realizar as previsões, e (2) aplicação do algoritmo DBSCAN para a criação dos conjuntos seguida da mesma rede LSTM para prever as variações de preço. O estudo foi realizado em um conjunto com 51 ações do mercado acionário brasileiro, e os experimentos sugeriram que utilizar um método para criar conjuntos de ações similares melhora os resultados em aproximadamente 7 porcento de acurácia e f1-score, e 8 porcento de recall e precision quando comparados a modelos para ações isoladas. / [en] Predicting stock market prices is a hard task. The main reason for that is due to the fact its environment is highly dynamic, intrinsically complex and chaotic. The traditional economic theories tell us that trying to predict short-term stock price movements is a wasted effort because the market is influenced by several external events and its behavior approximates a random walk. Recent studies, such as (37) and (51), address this problem and create specific prediction models for the price behavior of an isolated stock. This work presents a proposal to predict price movements based on stock sets considered similar. Our goal is building a model to identify whether the price tends to bullishness or bearishness in the (near) future, considering stock information from similar sets based on two sources of information: historical stock data and Google Trends news. Firstly, the proposed study applies a method to identify similar stock sets and then creates a predictive model based on LSTM (long short-term memory) for these sets. More specifically, two experiments were conducted: (1) using the K-Means algorithm to identify similar stock sets and then using a LSTM neural network to predict stock price movements for these stock sets; (2) using the DBSCAN algorithm to identify similar stock sets and then using the same LSTM neural network to forecast stock price movements. The study was conducted over 51 stocks of the brazilian stock market. The results suggested that using an algorithm to identify similar stock clusters yields an improvement of approximately 7 percent in accuracy and f1-score and 8 percent in recall and precision when compared to specific models for isolated stocks.

In this project, three primary problems were addressed to improve battery data management and software performance evaluation. All solutions used voltage values in time together with various device characteristics. Battery replacement labeling was performed using Hidden Markov Models. Both deep reinforcement learning, specifically TD3 with an LSTM layer, and traditional models were employed to predict future battery voltages. These predictions subsequently informed a developed novel method for early evaluation of software impact on battery performance. A baseline model was also introduced for optimal battery replacement timing. Results indicated that the TD3-LSTM model achieved a mean absolute percentage error below 5%, on par with traditional methods. The battery replacement labeling had above 85% correctly labeled replacements, impact on battery performance was above 90% correct in software comparisons. TD3-LSTM proved a viable choice for multi-step predictions requiring online learning, albeit requiring potentially more tuning. / I detta projekt behandlades tre primära problem i syfte att förbättra batteridatahantering och utvärdering av mjukvaruprestanda. Alla lösningar använde spänningsvärden i tid tillsammans med olika enhetsegenskaper. Batteribytesmärkning utfördes med hjälp av Hidden Markov Models. Både deep reinforcement learning, särskilt TD3 med ett LSTM-lager, och traditionella modeller användes för att förutsäga framtida batterispänningar. Dessa förutsägelser användes sedan i en framtagen ny metod för tidig utvärdering av mjukvarans påverkan på batteriprestanda. En basmodell introducerades också för optimal batteribytestid. Resultaten indikerade att TD3-LSTM modellen uppnådde ett genomsnittligt absolut procentfel under 5%, i nivå med traditionella metoder. Batteribytesmärkningen hade över 85% korrekt märkta batteribyten, inverkan på batteriprestanda var över 90% korrekt i mjukvarujämförelser. TD3-LSTM visade sig vara ett hållbart val för flerstegsförutsägelser som kräver onlineinlärning, även om det krävde potentiellt mer justering.

Integrating sophisticated statistical methods into database management systems is gaining more and more attention in research and industry in order to be able to cope with increasing data volume and increasing complexity of the analytical algorithms. One important statistical method is time series forecasting, which is crucial for decision making processes in many domains. The deep integration of time series forecasting offers additional advanced functionalities within a DBMS. More importantly, however, it allows for optimizations that improve the efficiency, consistency, and transparency of the overall forecasting process. To enable efficient integrated forecasting, we propose to enhance the traditional 3-layer ANSI/SPARC architecture of a DBMS with forecasting functionalities. This article gives a general overview of our proposed enhancements and presents how forecast queries can be processed using an example from the energy data management domain. We conclude with open research topics and challenges that arise in this area.