Mobile Edge Clouds (MECs) address the critical needs of bandwidth-intensive, latency-sensitive mobile applications by positioning computing and storage resources at the network’s edge in Edge Data Centers (EDCs). However, the diverse, dynamic nature of EDCs’ resource capacities and user mobility poses significant challenges for resource allocation and management. Efficient EDC operation requires accurate forecasting of computational load to ensure optimal scaling, service placement, and migration within the MEC infrastructure. This task is complicated by the temporal and spatial fluctuations of computational load.We develop a novel MEC computational demand forecasting method using Federated Learning (FL). Our approach leverages FL’s distributed processing to enhance data security and prediction accuracy within MEC infrastructure. By incorporating uncertainty bounds, we improve load scheduling robustness. Evaluations on a Tokyo dataset show significant improvements in forecast accuracy compared to traditional methods, with a 42.04% reduction in Mean Absolute Error (MAE) using LightGBM and a 34.93% improvement with CatBoost, while maintaining minimal networking overhead for model transmission. 

Load forecasting is a crucial topic in energy management systems (EMS) due to its vital role in optimizing energy scheduling and enabling more flexible and intelligent power grid systems. As a result, these systems allow power utility companies to respond promptly to demands in the electricity market. Deep learning (DL) models have been commonly employed in load forecasting problems supported by adaptation mechanisms to cope with the changing pattern of consumption by customers, known as concept drift. A drift magnitude threshold should be defined to design change detection methods to identify drifts. While the drift magnitude in load forecasting problems can vary significantly over time, existing literature often assumes a fixed drift magnitude threshold, which should be dynamically adjusted rather than fixed during system evolution. To address this gap, in this paper, we propose a dynamic drift-adaptive Long Short-Term Memory (DA-LSTM) framework that can improve the performance of load forecasting models without requiring a drift threshold setting. We integrate several strategies into the framework based on active and passive adaptation approaches. To evaluate DA-LSTM in real-life settings, we thoroughly analyze the proposed framework and deploy it in a real-world problem through a cloud-based environment. Efficiency is evaluated in terms of the prediction performance of each approach and computational cost. The experiments show performance improvements on multiple evaluation metrics achieved by our framework compared to baseline methods from the literature. Finally, we present a trade-off analysis between prediction performance and computational costs.

For efficient energy exchanges in smart energy grids under the presence of renewables, predictions of energy production and consumption are required. For robust energy scheduling, prediction of uncertainty bounds of Photovoltaic (PV) power production and consumption is essential. In this paper, we apply several Machine Learning (ML) models that can predict the power generation of PV and consumption of households in a smart energy grid, while also assessing the uncertainty of their predictions by providing quantile values as uncertainty bounds. We evaluate our algorithms on a dataset from Swedish households having PV installations and battery storage. Our findings reveal that a Mean Absolute Error (MAE) of 16.12W for power production and 16.34W for consumption for a residential installation can be achieved with uncertainty bounds having quantile loss values below 5W. Furthermore, we show that the accuracy of the ML models can be affected by the characteristics of the household being studied. Different households may have different data distributions, which can cause prediction models to perform poorly when applied to untrained households. However, our study found that models built directly for individual homes, even when trained with smaller datasets, offer the best outcomes. This suggests that the development of personalized ML models may be a promising avenue for improving the accuracy of predictions in the future.

Energy management within microgrids under the presence of large number of renewables such as photovoltaics is complicated due to uncertainties involved. Randomness in energy production and consumption make both the prediction and optimality of exchanges challenging. In this paper, we evaluate the impact of uncertainties on optimality of different robust energy exchange strategies. To address the problem, we present AIROBE, a data-driven system that uses machine-learning-based predictions of energy supply and demand as input to calculate robust energy exchange schedules using a multiband robust optimization approach to protect from deviations. AIROBE allows the decision maker to tradeoff robustness with stability of the system and energy costs. Our evaluation shows, how AIROBE can deal effectively with asymmetric deviations and how better prediction methods can reduce both the operational cost while at the same time may lead to increased operational stability of the system.

Prediction of solar power generation is important in order to optimize energy exchanges in future micro-grids that integrate a large amount of photovoltaics. However, an accurate prediction is difficult due to the uncertainty of weather phenomena that impact produced power. In this paper, we evaluate the impact of different clustering methods on the forecast accuracy for predicting hourly ahead solar power when using machine learning based prediction approaches trained on weather and generated power features. In particular, we compare clustering methods using clearness index and K-means clustering, where we use both euclidian distance and dynamic time-warping. For evaluating prediction accuracy, we develop and compare different prediction models for each of the clusters using production data from a swedish SmartGrid. We demonstrate that proper tuning of thresholds for the clearness index improves prediction accuracy by 20.19% but results in worse performance than using K-means with all weather features as input to the clustering.

Towards zero CO2 emissions society, large shares of renewable energy sources and storage systems are integrated into microgrids as part of the electrical grids for energy exchange aiming to effectively reduce the stress from the transmission grid. However, energy management within and across microgrids is complicated due to many uncertainties such as imprecise knowledge on energy production and demand, which makes energy optimization challenging. In this paper, we present an open architecture that uses machine learning algorithms at the edge to predict energy consumption and production for energy management in smart microgrids. Such predictions are aggregated across different prosumers at a centralized marketplace in the Cloud using Kafka Streams and OpenSource IoT platforms. Using pluggable optimization algorithms, different microgrids can implement different strategies for real-time optimal energy schedules. The proposed architecture is evaluated in terms of scalability and accuracy of predictions. Our heuristics can effectively optimize medium-sized microgrids.