Estimating CPU power on heterogeneous ARM-based commodity devices is challenging due to limited access to CPU's voltage domains. As a result, state-of-the-art energy-aware Federated Learning (FL) frameworks typically rely on simplified approximate power models to estimate computation energy, rather than the more accurate analytical CMOS-based model. To bridge this gap, we propose a reproducible CPU power estimation methodology combined with a rail-to-cluster mapping technique to retrieve cluster-level supply voltage. We evaluate our approach on two commodity Android devices and show that the analytical model predicts CPU power with errors below 10

Federated learning (FL) is a distributed machine learning (ML) scheme that enables collaborative model training without compromising data privacy. In the TinyML community, communication is often prioritized over computation due to the limited bandwidth of the client devices. This paper aims to explore the rationale behind this perception by focusing on small models with approximately 1M parameters and investigating the effects of quantization and sparsification techniques for computation and communication purposes. We introduce mathematical models essential for estimating the time and energy consumption of computation and communication in an FL setup. Our research explores various configurations that impact FL efficiency, including model quantization and sparsification, processor frequency, batch sizes, and ML architecture. Notably, we demonstrate that for small models, communication energy can be negligible compared to computational energy. Consequently, contrary to existing literature, we propose that sparsification for communication may not be necessary as it introduces additional computational effort and energy consumption to achieve a target accuracy without significant gains in communication energy. This research enhances our understanding of the trade-offs between communication and computation in FL, particularly for small models like in TinyML. By emphasizing the importance of computational energy, our findings provide valuable insights for the design and optimization of FL systems.

Electric vehicles (EVs) and their ecosystem have unquestionably made significant technological strides. Indeed, EVs have evolved into sophisticated computer systems with extensive internal and external communication capabilities. This interconnection raises concerns about security, privacy, and the expanding risk of cyber-attacks within the electric vehicle landscape. In particular, the charging infrastructure plays a crucial role in the electric mobility ecosystem. With the proliferation of charging points, new attack vectors are opened up for cybercriminals. The threat landscape targeting charging systems encompasses various types of attacks ranging from physical attacks to data breaches including customer information. In this paper, we aim to leverage the power of model-driven engineering to model and analyze EV charging systems at early stages. We employ domain-specific modeling language (DSML) techniques for the early security modeling and analysis of EV charging infrastructure. We accomplish this by integrating the established EMSA model for electric mobility, which encapsulates all key stakeholders in the ecosystem. To our knowledge, this represents the first instance in the literature of applying DSML within the electric mobility ecosystem, highlighting its innovative nature. Moreover, as our formalization based on DSML is an iterative, continuous, and evolving process, this approach guarantees that our proposed framework adeptly tackles the evolving cyber threats confronting the EV industry. Specifically, we use the Object Constraint Language (OCL) for precise specification and verification of security threats as properties of a modeled system. To validate our framework, we explore a set of representative threats targeting EV charging systems from real-world scenarios. To the best of our knowledge, this is the first attempt to provide a comprehensive security modeling framework for the electric mobility ecosystem.

Energy efficiency is a critical concern for machine learning (ML) algorithms deployed in data centers. Recently, many works in the literature have focused on running ML algorithms on energy-efficient and constrained hardware, such as mobile phones, to reduce the energy footprint of training ML models. This paper introduces the Power Profiler, an open-source monitoring platform that provides valuable insights into the energy consumption of ML algorithms on Android mobile devices. By capturing key performance indicators (KPIs) such as voltage, current, and CPU usage, the Power Profiler enables real-time monitoring of their energy usage. It eliminates the need for custom hardware installations and facilitates the development of energy-efficient ML models. The Power Profiler can empower researchers to understand and optimize the energy consumption patterns of ML algorithms, facilitating the creation of sustainable ML models for energy-efficient mobile deployments.

Renewables are the greener substitute for the conventional polluting sources of generating energy. For their successful integration into the power grid, accurate forecasts are required. In this paper, we report the lessons acquired from our previous works on generating time-series ARIMA-based forecasting models for renewables. To this end, we considered a consistent dataset spanning the last four years. Assuming four different performance metrics for each of the best ARIMA-based models of our previous works, we derived a new optimal model for each month of the year, as well as for the two different methodologies suggested in those works. We then evaluated the performance of those models, by comparing the two methodologies: in doing so, we proposed a hybrid methodology that took the best models out of those two methodologies. We show that our proposed hybrid methodology has improved yearly accuracy of about 89.5% averaged over 12 months of the year. Also, we illustrate in detail for the four years under study and each month of the year the observed percentage of renewables and its corresponding accuracy compared to the generated forecasts. Finally, we give the implementation details of our open-source REN4KAST software platform, which provides several services related to renewables in Germany.

Data centers can participate in demand-response schemes by reducing their demand, however, at the expense of the agreed-upon performance of their IT services defined by the SLAs. The successful application of such schemes necessitates a careful analysis so that the amount of degradation of the SLAs with respect to power savings can be quantified helping the data center operators to set up the optimal configuration. In this paper, we study and analyze a system consisting of a data center, its operator, and IT clients under the consideration of relaxed SLAs. For this purpose, we consider a data center system consisting of two heterogeneous pools of servers, where each server is modeled using the single-server system with a power-saving inactive state, non-zero (random) activation/deactivation times, and hot standby state. Making use of the distributional Little's Law, derive the steady-state performance (in terms of response time distribution) and average power demand and study the power-performance trade-off in an explicit way. Numerical results illustrate the model's theoretical properties, under different considerations of low, medium, and high workload utilization rates.

The integration of smart thermostats in home automation systems has created an opportunity to optimize space heating and cooling through the use of machine learning, for example for thermal model identification. Nonetheless, its full potential remains untapped due to the lack of a suitable learning scheme. Traditional centralized learning (CL) and federated learning (FL) schemes could pose privacy and security concerns, and result in a generic model that does not adequately represent thermal requirements and characteristics of each individual home. To overcome these limitations, in this paper we embrace the novel peer-to-peer learning scheme for on-device training of home thermal models. Specifically, we adapt the personalized peer-to-peer algorithm proposed in recent work (called P3) to efficiently train personalized thermal models on resource-constrained devices. Our preliminary experiments with data from 1,000 homes, using the LSTM model, demonstrate that the adapted P3 algorithm produces accurate and personalized thermal models while being extremely energy-efficient, consuming respectively 600 and 40 times less energy than the CL and FL schemes. This result suggests that the P3 algorithm offers a privacy-conscious, accurate, and energy-efficient solution for training thermal models for the many homes in the building stock.

Smart thermostats are increasingly popular in homes and buildings as they improve occupant comfort, lower energy use in heating and cooling systems, and reduce utility bills by automatically adjusting room temperature according to measurements of their built-in sensors. To maximize energy savings, many of these thermostats use machine learning (ML) for more accurate prediction and optimal control. These ML models must be trained for each individual building to ensure higher thermal comfort and energy savings, owing to the fact that buildings are custom-built, can be located in different climates, and may have unique occupancy patterns. This kind of training requires a significant amount of energy and sharing raw data emitted by the sensors in each building. To address these issues, we propose a novel methodology to train accurate and personalized thermal models for each home with a minimal energy footprint. Specifically, we use temporal and spatial abstraction to downsample sensor data and cluster homes with similar characteristics to train representative thermal models for each cluster. These models are customized for each home using meta-learning, achieving personalized models with high accuracy. Additionally, combining multi-step ahead prediction with our proposed abstraction technique would permit resource-constrained devices (e.g., microcontrollers) to accurately forecast indoor temperature for larger time intervals with negligible computation overhead. Experiments with smart thermostat data from 1,000 homes show that our methodology offers accurate and personalized thermal models with substantial savings in network bandwidth and training energy. To be specific, our methodology is approximately 384 (300) times more energy efficient than training an LSTM (RNN) model using the conventional approaches. Given the anticipated increase in the demand for smart thermostats and the fact that thermal models must be (re)trained regularly (e.g., every season), the proposed methodology could significantly reduce the environmental impact of training ML models for thermal comfort optimization in the long run.

Nowadays, the integration of home automation systems with smart thermostats is a common trend, designed to enhance resident comfort and conserve energy. The introduction of smart thermostats that can run machine learning algorithms has opened the door for on-device training, enabling customized thermal experiences in homes. However, leveraging the flexibility offered by on-device learning has been hindered by the absence of a tailored learning scheme that allows for accurate on-device training of thermal models. Traditional centralized learning (CL) and federated learning (FL) schemes rely on a central server that controls the learning experience, compromising the home’s privacy and requiring significant energy to operate. To address these challenges, we propose PePTM, a personalized peer-to-peer thermal modeling algorithm that generates tailored thermal models for each home, offering a controlled learning experience with a minimal training energy footprint while preserving the home’s privacy, an aspect difficult to achieve in both CL and FL. PePTM consists of local and collaborative learning phases that enable each home to train its thermal model and collaboratively improve it with a set of similar homes in a peer-to-peer fashion. To showcase the effectiveness of PePTM, we use a year’s worth of data from US homes to train thermal models using the RNN time-series model and compare the data across three learning schemes: CL, FL, and PePTM, in terms of model performance and the training energy footprint. Our experimental results show that PePTM is significantly energy-efficient, requiring 695 and 40 times less training energy than CL and FL, respectively, while maintaining comparable performance. We believe that PePTM sets the stage for new avenues for on-device thermal model training, providing a personalized thermal experience with reduced energy consumption and enhanced privacy.

In data centres, there exist several techniques for energy efficiency purposes. When applied, most of those techniques have impact on the quality (e.g. performance) of the underlying services. A careful study is required in order to optimise such an energy-performance trade-off. In this paper, we study the speed scaling as an energy efficiency technique within the scope of high-performance computing (HPC) data centres. We propose a methodology based on three-levels: analytic, simulation and technical. At the analytical level, the matrix-analytic method (MAM) allows one to obtain energy-performance measures explicitly for a small-scale system. At the simulation level, discrete-event simulation (DES) based on the generalised semi-Markov processes (GSMP) is used to derive the corresponding estimates. Finally, at the technical level, a real small-scaled system in a controlled environment is used. The preliminary results demonstrate that simulation and technical models go well together with the theoretical one with an accuracy of more than 95%.

As electric vehicles (EVs) are slowly becoming a common occurrence on roads, commercial EV charging is becoming a standard commercial service. With this development, charging station operators are looking for ways to make their charging services more profitable or allocate the available resources optimally. Dynamic pricing is a proven technique to increase revenue in markets with heterogeneous demand. This paper proposes a Markov Decision Process (MDP)-based approach to revenue- or utilization- maximizing dynamic pricing for charging station operators. We implement the method using a Monte Carlo Tree Search (MCTS) algorithm and evaluate it in simulation using a range of problem instances based on a real-world dataset of EV charging sessions. We show that our approach provides near-optimal pricing decisions in milliseconds for large-scale problems, significantly increasing revenue or utilization over the flat-rate baseline under a range of parameters.