Soiling is the accumulation of dirt in solar panels that leads to a decreasing trend in solar energy yield and may be the cause of vast revenue losses. The effect of soiling can be reduced by washing the panels, which is, however, a procedure of non-negligible cost. Moreover, soiling monitoring systems are often unreliable or very costly. We study the problem of estimating the soiling ratio in photovoltaic (PV) modules, i.e., the ratio of the real power output to the power output that would be produced if solar panels were clean. A key advantage of our algorithms is that they estimate soiling, without needing to train on labeled data, i.e., periods of explicitly monitoring the soiling in each park, and without relying on generic analytical formulas that do not take into account the peculiarities of each installation. We consider as input a time series comprising a minimum set of measurements that are available to most PV park operators. Our experimental evaluation shows that we significantly outperform current state-of-the-art methods for estimating soiling ratio.
More and more aspects of our everyday lives are influenced by automated decisions made by systems that statistically analyze traces of our activities. It is thus natural to question whether such systems are trustworthy, particularly given the opaqueness and complexity of their internal workings. In this paper, we present our ongoing work towards a framework that aims to increase trust in machine-generated recommendations by combining ideas from three separate recent research directions, namely explainability, fairness and user interactive visualization. The goal is to enable different stakeholders, with potentially varying levels of background and diverse needs, to query, understand, and fix sources of distrust.
Telco edge data centers (DCs) accommodate applications whose load fluctuates considerably during a day. This variability mandates swift and responsive resource adjustments to mitigate the risk of incurring unnecessary costs during off-peak periods, where a significant fraction of nodes may be under-utilized. Tackling this challenge is integral to optimizing operational efficiency and cost-effectiveness of telco edge DCs. To this end, this article aims to address the Dynamic Data Center Sizing (DDS) problem, which boils down to optimizing the number of active nodes as per current resource demand. The proposed DDS solution consists of two core modules, namely a forecasting module, which predicts resource demands, and a decision-making module, which acts upon predicted demands. The decision-making of DDS is implemented via the filtering and the rank-drain-observe (RDO) algorithms. Filtering is based on integer linear programming, and it computes the theoretically optimal state of the DC based on predicted resource demands. RDO is a heuristic that strives to realize the optimal DC state in an iterative and robust fashion. To expedite RDO in large-scale clusters, we further devise a Reinforcement Learning (RL)-enabled DDS variant (i.e., RL-DDS), in which RDO integrates an RL agent that computes optimized batches of nodes for concurrent deactivation. We propose an innovative solution based on the notion of Digital Twin to train the RL agent in emulation mode. DDS and RL-DDS are evaluated upon real-life testbeds resembling actual DCs. Results demonstrate significant cost reduction, ranging from 7% in conservative and relatively static scenarios, up to 38% in highly dynamic settings.