Artificial Intelligence Techniques for a Scalable Energy Transition

This chapter focuses specifically on the development of a smart building energy management (SBEM) system. The system has two main goals: the first one is, by using an artificial neural network, to estimate and predict the thermal behavior of a large-scale building including instrumented and non-instrumented thermal zone. The second one consists, via a human graphical interface, in providing different advice to users to educate and attract them about energy reduction challenges. The originality of this chapter is close to the nature of the intervention of the system. It does not act directly on the HVAC building systems, as an automation system could realize, but on the USER. In other terms, it consists in inviting USER to have a good behavior to reduce energy consumption. So, by means of the student residential located in Douai in the North of the France, we validate the thermal behavior model developed and we also realize different factors analysis which may affect the energy consumption for optimization purposes. This leads in setting well the human interface to be sure that each user sticks to each advice in order to guarantee an efficient smart building energy management system design.

The built environment is responsible for more than a third of total energy consumption and greenhouse gas emissions in both Europe and North America. Electrification of heat and transport, as well as decarbonization through efficiency improvements and distributed energy resources is paving the way for a more sustainable built environment. Today, in many parts of the world, favourable policy regimes and technological advances are further accelerating this transition. However, these rapid changes have also spawned a number of issues, most notably in the form of increased electricity usage, new peak loads, and reverse power flows into the grid. These changes affect not just the distribution grid, but also the transmission grid through increased demand and steep ramp rate requirements due to the intermittency of renewable energy sources. Demand side management algorithms, often powered by the latest advances in artificial intelligence, offer a potential solution to this problem. However, these solutions are marred by data and computational requirements, as well as privacy concerns. Transfer learning has recently been shown to help avoid the requirement of copious amounts of data required to learn a model necessary for optimization. Likewise, federated learning is one potential solution to addressing user privacy concerns by learning from data in a distributed manner. Finally, reinforcement learning can do away with a number of lingering issues in classical model predictive control, especially enabling services which require fast response times such as frequency regulation. These advances cover the entire spectrum of data-driven demand side management offering, which will form the basis for not just more sustainable buildings but also a smarter energy grid in the future.

Smart grids present a paradigm-shift in which utilities and consumers participate in a bilateral communication, enabling the demand side to offer flexibility and electricity. Benefits from this transition range from improved power distribution to reduced dependency on the system. Buildings hold an essential role in the success of the paradigm. They need to be able to adopt flexible consumption patterns and to support the demands from the system. The objective of the work in this chapter is to render buildings ready to provide their flexibility and to participate in demand-response schemes. We propose a multi-agent energy management system, based on the Alternating Direction Method of Multipliers, which optimises power consumption and injection and supports demand-response requests from external parties, while ensuring efficiency, scalability, and privacy. Various experiments were conducted to validate the proposal. The results show significant energy cost savings and prove the feasibility of adopting various demand-response programs.

Residential energy smart management (RESM) has received considerable momentum in the recent decade considering its strong impact on the total energy consumption and the elaboration of the smart grid. Non-Intrusive Load Monitoring (NILM) is the first brick of the smart grid. In this paper, the importance of NILM in the smart grid is highlighted and its impact on different smart grid issues is discussed. Challenges facing NILM are also explained and existing solutions are reviewed. Mainly, an overview of different machine learning approaches is presented and these methods’ limits are discussed giving rise to open problems in the state of the art.

The chapter proposes neural networks and statistical decision making for fault diagnosis in energy conversion systems. It considers the condition monitoring problem for an energy conversion system comprising a solar power unit, a DC-DC converter, and a DC motor. The dynamic model of this energy conversion system is taken to be unknown and is reconstructed from its input and output measurements, being accumulated at different operating conditions, and taking finally the form of a neural network. Actually, the neural model consists of a hidden layer of Gauss–Hermite polynomial activation functions and an output layer of linear weights. The neural network is trained with the use of first-order gradient algorithms and the resulting model is taken to represent the fault-free functioning of the energy conversion system. To conclude about the existence of a fault, the measurements of the real output of the energy conversion system are compared against the estimated outputs which are provided by the neural model. Thus, the residuals’ sequence is generated. It is shown that the sum of the squares of the residuals’ vectors, multiplied with the inverse of the associated covariance matrix, stands for a stochastic variable (statistical test) which follows the χ ² distribution. By selecting the 96% or the 98% confidence intervals of this distribution one can have a precise and almost infallible decision making tool about the appearance of faults in the energy conversion system.

In this chapter, a hybrid method for fault diagnosis and prognosis, based on the notion of similarity measurement and similarity speed calculation, is proposed and applied to a wind turbine. The causal and structural properties of the physical model are used to identify the measured variables that carry the degradation process to be used as health indicators. The physical model is then used to generate data in normal, faulty, and failure operations and to identify failure thresholds. The fault diagnosis step is based on a multi-class support vector machine classification of attributes extracted from current measurement of wind turbine generator. Once a fault is detected and located, the fault prognosis module is triggered in order to estimate the remaining useful lifetime before observing wind turbine failure. To overcome the nonexistence of knowledge about the degradation trend, a geometric method based on Euclid metric is used for RUL estimation. The obtained results, evaluated using universal metrics, show the effectiveness and accuracy of the proposed method.

Scientific research in the area of fault prognosis is increasingly focused on estimating the Remaining Useful Life of equipment, since its knowledge is a key input to the scheduling of Condition-Based and Predictive Maintenance. Several research studies have been directed to developing methods for modeling the trend of health indicators for Remaining Useful Life estimation, this paper makes a review of these approaches. Fault diagnosis methods sensitive to the progressive evolution of degradation phenomena are presented and their usability for fault prognosis is discussed. Then, methods for modeling the trends of health indicators are analyzed to highlight the selection criteria of the modeling methods, according to the available information on the operating conditions of the systems, and on the degradation phenomenon. Finally, some reflections are made regarding the elements that prevent the large-scale use of prognostics in industry today, and on the integration of prognostics in risk assessment and management.

This chapter addresses the application of machine learning algorithms to detect attacks against smart grids. Smart grids are the result of a long process of transformation that power systems have been through, relying on Information and Communication Technology (ICT) to improve their monitoring and control. Although an objective of this convergence of power systems and ICT is to increase their reliability, the dependency on information technology has brought new cybersecurity vulnerabilities to this scenario. Therefore, developing new cybersecurity measures for smart grids is a key factor in their success. One of these measures is attack detection, which allows the timely mitigation of attacks with the aim of limiting possible damages to the targets. As machine learning algorithms have been widely applied as powerful tools to support the design of cybersecurity solutions in multiple areas, they also have huge potential for addressing the new challenges that smart grids pose. With this as the foundational perspective, this study starts by presenting an overview of smart grids, followed by possible attacks. After this discussion, we examine the background concepts for attack detection and machine learning. Then, we discuss the existing solutions, showing in detail how they address the particularities of smart grids and their attack types using machine learning algorithms. This is supplemented by a discussion of the open issues in the use of machine learning for smart grid attack detection, followed by some future research directions.

Artificial intelligence is anticipated to play a significant role in the smart homes of the future. Decisions have to be made based on a variety of information that will be available to the home occupants. With regard to electricity consumption, it is expected that price directed markets will allow home occupants to become price receivers at a resolution of very short-term intervals of time—prices may be sent in intervals of a few seconds. In that time frame, decision patterns cannot be formed with the physical participation of the home occupants. To fill this gap, artificial intelligence offers the necessary tools to develop smart decision-making algorithms that make automated efficient decisions. In this chapter, a new approach for making decisions with regard to electricity consumption of smart appliances is presented. In particular, a neurofuzzy anticipatory approach—that integrates neural networks with fuzzy inference—is presented as a means to make decisions over the length of the operational time of a smart appliance. The goal of the approach is to utilize the current operational variables values and price information together with their future projections to make decisions over the operational time interval of a smart appliance. The determination of the operational time of each appliance, when aggregated implicitly shapes the demand response of the occupant in the price directed market. The proposed neurofuzzy approach is tested on a set of simulated data from an HVAC system obtained with the GridLAB-D simulation software, and real world price signals.

Keeping the power supply of autonomous and electrical vehicles working even in case of faults is of uttermost importance in order to maintain the desired behavior during operation. Especially in case of increased autonomy faults occurring in the power supply when driving should not require the vehicle to stop operation immediately. Instead the autonomous vehicle should still be able to reach a safe state like an emergency lane or a parking space. In this chapter, we introduce a method that enables the development of battery systems that react on internal or external faults in a smart way. In particular, we discuss model-based reasoning for this purpose and show its application for configuring and diagnosing systems. Besides discussing the foundations behind model-based reasoning, we make use of a smart battery system as a case study. In addition, we describe how to use the corresponding physical model for fault detection and a logical model for computing the root cause of the observed failure. The intention behind the chapter is to provide all necessary details of the methods allowing to adapt the methods to implement similar smart adaptive systems.

The potential of distributed energy resources (including, flexible electrical loads) in providing grid services can be maximized with the recent advancements in demand side control. Effective coordination of the flexible loads for grid services, while satisfying end-user preferences and constraints, requires the knowledge of aggregated predictive flexibility of the distributed energy resources (DERs). Recent works have shown that the aggregated predictive flexibility of DERs can be modeled as a virtual battery (VB) whose state evolution is governed by a first-order dynamics including self-dissipation rate and energy capacity. Identifying the VB model parameters for a collection of DERs, however, is challenging primarily due to the following reasons: (1) the composition of DERs is time-varying and uncertain, with the device availability determined by uncertain end-user behavior, (2) the underlying device models and parameters are mostly unknown and uncertain, and (3) lack of available behind-the-meter sensing and measurements (partly due privacy concerns). As such, data-driven deep learning based frameworks have been proposed in this work to identify aggregated predictive flexibility models of a collection of DERs, using front-of-the-meter data (such as net power consumption, etc.). The effectiveness of the proposed frameworks is demonstrated on an ensemble of residential air conditioners and electric water heaters.

In this chapter, we introduce various applications for artificial neural networks in the context of power systems. Due to a fast pace of development in recent years, multiple libraries for setting up and training artificial neural networks are available as open-source software. In the field of power system analysis, the open-source software pandapower enables broad-scale automation of power flow calculations. Based on these developments, we present multiple applications for grid planners and grid operators that are based on supervised learning. The first application is the approximation of power flows, including line contingencies, in annual time series simulations. It enables grid planners to detect violations of operational constraints quickly. Secondly, a monitoring method trained on a yearly time series uses a low number of measurements to deliver real-time insights into the grid’s state to grid operators. Similarly, grid operators can use artificial neural networks for building grid equivalents that provide information about external grids under dynamic conditions. Lastly, artificial neural networks have proven well-suited to determine grid loss as a function of topological features like line length, distributed generation, etc.