Intelligent System Designs
Data-driven Sensor Calibration & Smart Meter Privacy
Time: Fri 2022-08-26 09.00
Location: Kollegiesalen, Brinellvägen 8, Stockholm
Video link: https://kth-se.zoom.us/j/68814277252
Subject area: Electrical Engineering
Doctoral student: Yang You , Teknisk informationsvetenskap
Opponent: Assistant professor Olga Fink, ETH Zürich, Switzerland
Supervisor: Professor Tobias J. Oechtering, Teknisk informationsvetenskap
Nowadays, the intelligent system has gained high popularity in successful implementation of real-time tasks due to its capability of providing efficient and powerful decision making in real applications. In this thesis, we aim for exploring and exploiting different concepts or methods to handle different tasks towards the intelligent system design. In particular, we focus on the following two problems: (i) Consumer-centric privacy-cost trade-off in smart metering system; (ii) Data-driven calibration for gas sensing system.
For the first target problem, an optimal privacy-preserving and cost-efficient energy management strategy is designed for each smart grid consumer that is equipped with a rechargeable energy storage. The Kullback-Leibler divergence rate is used as privacy measure and the expected cost-saving rate is used as utility measure. The corresponding energy management strategy is designed by optimizing a weighted sum of both privacy and cost measures over a finite time horizon, which is achieved by formulating our problem into a partial observed Markov decision process problem. A computationally efficient approximated Q-learning method is proposed as a extension to high-dimensional problems over an infinite time horizon.
Furthermore, the privacy-preserving and cost-efficient energy management strategy is designed for multiple smart grid consumers that are equipped with renewable energy sources. Different from the previous problem, the adversary is assumed to employ a factorial hidden Markov model based inference for load disaggregation, and the corresponding joint log-likelihood of the model is utilized as privacy measure. A dynamic pricing model is studied, where the price of unit amount of energy is determined by the consumers' aggregated power request, which suits a commodity-limited market. The consumers' energy management strategy is designed under a non-cooperative game framework by optimizing a weighted sum of both privacy measure and the user's energy cost savings. The consumers' non-cooperative game is shown to admit a unique pure strategy Nash equilibrium. As an extension, a computational-efficient distributed Nash equilibrium energy management strategy seeking method is proposed, which also avoids the privacy leakage due to the sharing of payoff functions between consumers.
For the second target problem, several data-driven self-calibration algorithms are developed for low-cost non-dispersive infrared sensors. The measurement errors of the sensors are mainly caused by the remaining model errors and can be fully described by the drift of the calibration parameter. This leads to our first formulation of a statistical inference problem on the true calibration parameter under the HMM framework, which is a stochastic model that jointly builds on different quantities introduced by the physical model. To better track the time-varying drift process of the sensor, a time-adaptive expectation maximization learning framework is proposed to efficiently update the HMM parameters. For the joint calibration of the gas sensing system, sensors firsttransmit their belief functions of the true gas concentration levelto the cloud. Then the cloud fusion center computes a fusedbelief function according to certain rules. This belief functionis then used as reference for calibrating the sensors. To dealwith the case where belief functions highly conflict with eachother, a Wasserstein distance based weighted average belieffunction fusion approach is first proposed as networked calibration algorithm. To achieve more long-term stable calibration results, the networked calibration problem is further formulated as a partially observed Markov decision process problem, and the calibration strategies are derived in a sequential manner. Correspondingly, the deep Q-network approach is applied as a computationally efficient method to solve the proposed Markovdecision process problem.
The results in this thesis have shown that our proposed design frameworks can provide concise but precise mathematical models, proper problem formulations, and efficient solutions for the target design objectives of different intelligent systems.