Professor Kang Li

 

Smart Energy Systems Group

Institute of Communication and Power Networks

University of Leeds

 

李慷

英国利兹大学电子电器工程学院

智慧能源系统教授

通讯及电力网络研究所所长

Project 1. Engineering genes:

An AI orchestrated physics informed data driven approach for modelling complex systems

We first investigated a subset selection problem in the following modelling paradigm. Suppose variable (or feature) y can be represented by equation (1): y=f(g_1, g_2,…, g_m)+e, where f is a model structure or a network configuration to fuse a set of low-dimensional functions g_i, i=1,…,m, and e is the modelling residual. If f is in the form of linear-in-the-parameter structure, and a large number of g_i (e.g. basis functions, model terms, etc) may be possibly fused into this model structure, then the aim is to select a much smaller set of g_i to approximate y, given a criterion or a set of criteria. This is a subset selection problem which is often computational very expensive if the number of low-dimensional functions g_i is large. We first proposed a fast forward selection algorithm (FRA) in [1] in 2005, which works directly on the sum squared errors by introducing a recursive residual matrix, in contrast to the popular Orthogonal Least Squares (OLS) method which applies orthogonalization directly on the regressors. This allows faster and more stable forward regressor selection. This framework is further extended to enable a fast and effective two-stage procedure where the second stage removes or replaces redundant regressors obtained in the first stage of the forward subset selection, leading to a more compact and locally optimal model as detailed in [2] in 2006. The mathematical framework of the two-stage approach was then extended to the OLS, resulting in a more efficient two-stage OLS method [3]. The framework is further extended to a few popular model types that have been intensively used in the machine-learning and artificial intelligence field, such as RBF networks, Single Layer Forward Neural Networks, [4]-[8], fuzzy neural networks, least squares support vector machines (LS-SVM), and nonlinear principal component analysis, etc [9]-[13], where g_i has tunable parameters. These methods have been applied to system modelling and identification, system monitoring and fault diagnosis, and nonlinear control. Internal collaborations in the identification and machine learning fields include Wiener-Hammerstein modelling [14]-[[16], non-parametric modelling [17], imbalanced data and fusion of heterogeneous temporal spatial data [18], classification of large data-sets [19], and deep reinforcement learning [20], etc. A survey paper on nonlinear system identification was highly commended by the International Journal of Systems Science [21].

During the same period, we consider a large class of complex processes which are governed by a set of first principle and empirical laws spanning across different time and length-scales, while these ‘a priori’ knowledge are often formulated as high dimensional nonlinear partial difference equations which are computationally extremely expensive to solve. Based on the Kolmogorov and Duhamel superposition theorems for multidimensional problems, these equations can be represented by a combination of low dimensional functions as formulated in equation (1). The challenges include the identification of model structure f in (1) and the basis functions g_i. To address these challenges, we have proposed the eng-genes modelling framework in 2005 [22], which in (1) the low dimensional functions g_i are selected among a group of salient fundamental functions (engineering genes) taken from the first principle laws governing the underlying physics for the systems under study, while the structure f in (1) is identified by using a combination of simple operations (multiplication and superposition) on the genes to produce transparent/interpretable model to replace the original complex systems. To produce such an eng-genes model for a complex system, data-driven bio-inspired heuristic optimization algorithms have also been investigated. The eng-genes idea has been applied to the modelling of both engineering systems as well as biological systems [23][24][25]. The engineering genes framework shares a lot of commonalities with a range of recently proposed concepts in machine learning, such as transfer learning, ensemble learning, physics-informed machine learning/neural network/ deep networks, etc. Our recent interests are to investigate multi-scale problems across temporal and spatial dimensions.

This research was primarily funded by EPSRC under GR/S85191/01 in 2004, and engineering applications have been funded by a few follow-on projects from both research councils and industry.

Project 2. Energy storage systems and smart grid

My research on energy and power started in 1998 since I moved to Queen’s University Belfast as a research fellow, researching NOx emission modelling in thermal power plants in first few years [26]. My focus shifted to smart grid and energy storage systems since 2009 primarily through RCUK Science Bridge project under grant EP/G042594/1 and 2013 EPSRC funded iGIVE project under grant EP/L001063/1. My research covers unit commitment and economic load dispatch considering renewable generations and EV charging [27]-[31], wide area power system monitoring and fault detection using PMUs [32][33], stochastic modelling and forecasting of wind power [34][35], control of multi-terminal HVDC systems for integrating offshore wind farms [36][37], multi-vector energy systems and energy market [38]-[42], microgrid and district energy management [43]-[45], wireless EV charging [46], waste heat recovery using organic Rankine Cycle [47][48], and management of battery energy storage systems for electric vehicles and power grid. We have developed an advanced laboratory on the smart grid and EVs, with up to date battery testing, characterization and grid interfaced charging/discharging facilities at different operation and ambient conditions.

Battery energy storage is a key technology for enabling a greater share of renewable energy in the power sector and for transportation electrification. The battery management system plays a critical role in delivering safe, reliable and effective operation across cell, modular and package levels. We developed a series of novel modelling, monitoring, control and optimization technologies for battery systems and their applications in electric vehicles and smart grid, in collaboration with industrial and academic partners. We first proposed an integrated approach for real-time model-based SOC estimation of lithium-ion batteries to address flat OCV vs SOC relationship and the hysteresis effect characterizing some of the Li-ion batteries like LFPs [49]. Considering that temperature and temperature uniformity have significant impacts on the battery safety, performance, and lifespan, hence the importance of real-time battery internal and external temperature profile monitoring can never be overestimated. We therefore developed a simplified battery thermoelectric model structure for battery thermal and electrical modelling and online internal state estimation [50][51]. With these models at hand, we developed optimal charging controls for EV batteries, incorporating charging constraints on the internal temperature rising, charging time and ageing factor [52]-[55]. Considering that low cost flow batteries will become increasingly important in the power sector, in particular for off-grid and mini-grid applications in developing countries, we have collaborated with Fraunhofer ICT and City University in developing a 25Ah Zinc-Nickel single flow battery demonstrator, and we have investigated the corresponding Zn-Ni flow battery management and monitoring method [56]-[58]. Our most recent focus is on the development of FBG optical fibre sensing system and associated AI orchestrated data driven analytic platform for battery design, degradation analysis, and real-time key state monitoring and thermal management [59] and machine-learning based SOH estimation of batteries [60], which have a great potential for significantly improving the safety and life-span of battery storage systems for future high-power application scenarios in power grid and transportation electrification.

Project 3. Energy management for industry decarbonization and Industry 4.0

I have a deep-rooted interest in industry decarbonization due to my control engineering and informatics background, and I believe our next generation of zero-carbon industry holds the promise for future wealth generation, for both developed and developing countries. My research in this field in the past decade started from the EPSRC projects EP/F021070/1, and the activities have been funded by various follow-on projects from both research councils and industry.

We started our research on polymer processing first. The polymer industry makes a major contribution to the UK economy, and extrusion is a fundamental polymer processing stage. However, considerable waste of both energy and raw materials can occur during lengthy change-over and start-up periods, and conservative operating conditions make this situation worse. Furthermore, the extrusion process is highly complex, and incomplete melting of the polymer may occur at inappropriate processing conditions that give rise to poor mixing and localised hot and cold spots, which in turn lead to pulsations in the throughput. As the key melt indicator, viscosity is difficult to monitor online, and as a result of this and the process characteristics described it is difficult to control the quality of an extrudate material and hence the final product. In this regard, we have made considerable progress on monitoring and control of polymer extrusion to reduce set-up times, waste and energy consumption and improve product quality: 1) We have investigated the pressure model incorporating changes in the machine/die and material properties [61]. 2) We have developed a model identification procedure based on grey-box modelling with a genetic algorithm for viscosity modelling, and further improved a soft sensor approach with a feedback structure to enable the online adaptability of the viscosity model in response to modelling errors and disturbances, hence producing a reliable estimate of viscosity [62][63]. 3) We have developed a model free fuzzy controller to minimize the variation in melt pressure and temperature, thus reducing the variation of viscosity [64]. 4) We have developed torque signal acquisition methods and investigated an inferential monitoring approach to the screw load torque signal in an extruder [65]. 5) We have employed a thermocouple mesh technique developed by the University of Bradford to measure the die melt temperature profile of a single screw extruder and developed a nonlinear model to predict the die melt temperature profile from readily measured process parameters. The model was then used to identify the effects of individual processing parameters on the die melt flow homogeneity and to further identify the optimal process settings to minimise melt temperature variance across the melt flow [66][67][68]. 6) We have explored energy constraints and developed a strategy for responding to process instabilities, then investigated a model based fuzzy control approach to maintain the die melt temperature variance across the melt flow while achieving the desired average die melt temperature. We have further developed a fuzzy controller for the control of the polymer melt quality including viscosity [69]. 7) We have developed a novel nonlinear multivariate statistical approach based on neural networks for fault detection in polymer processing [70]. These methods have also been applied to the modelling and control of injection stretch blow moulding [71]. All these new developments have considered in the industrial environment, and close collaboration with other leading research groups in the UK and with local industrial companies has enabled the direct application of the techniques developed to maximize the impact. Detailed report can be found at here.

As an outcome of this research, we have developed a minimal-invasive low-cost cloud based energy monitoring and analytic platform (Point Energy Technology) used in several industrial sectors, including polymer processing and food processing, winning several prestigious awards, including InstMC ICI prize 2015, INVENT 2016 award, finalist of Sustainable Energy Awards 2016 by Sustainable Energy Authority of Ireland, and Outstanding Award from Knowledge Transfer Partnerships 2015. The platform is able to offer low-cost component level energy monitoring, providing actionable insights that reduce the energy costs and optimise the manufacturing processes. Hence it brings tangible benefits on reduced energy consumption, reduced maintenance, reduce machine downtime, increased productivity and increased product quality.

As a follow-on application, we have deployed the Point Energy platform in the bakery industry, leading to a few tangible results. 1) Statistical analysis of electricity consumption data over a seven-day period is conducted for a local bakery, including the identification of operational modes for individual processing units using an enhanced clustering method and the voltage unbalance conditions associated with these identified modes. Two technical strategies, namely electrical load allotment and voltage unbalance minimisation, are then proposed, which could attain more than 800 kwh energy saving during this period and the current unbalance could be reduced to less than 10% [72]; 2) We used AI technique to solve the job shop scheduling problem based upon the commercial electricity tariffs, and this reduces the electricity bill by £80 per day in the case study [72]. 3) We have introduced a F-QoS metric which is able to classify the quality of the data stream from a WSN, and applied to a LoRaWAN network we used in our Point Energy platform deployed in a large commercial bakery with a low-disruption installation, where the network links are strained by large metal obstructions and the endpoints are installed inside metal cabinets. The results indicate that the LoRaWAN deployed in our system is capable of data acquisition in an unadapted and challenging environment, with the recommendation that the raw sample rate should be triple the desired final sample rate [73]. 4) We have developed an orthogonal-least-square-based autoencoder to generate new samples for the imputation of missing values in the data collected from harsh industrial environment, and it outperforms significantly alternative approaches while the missing ratio is greater than 0.05 [74].

Our next step is to roll out the Point Energy platform and associated AI technologies for more SMEs during their low-carbon or decarbonization journey, embracing significant amount of renewable energy in combination with our developed energy storage technologies in Project 2.

Project 4. Transport decarbonization

Transportation decarbonization is crucual in achieving global decarbonization goals. Electrified transportation means (road, marine, aviation and rail) powered by renewable sources will not only bring substantial environmental benefits on top of social and economic benefits, but also create a few technical challenges. Energy storage, power electronics, ICT, AI and big data analytics are among key enabling technologies to make it happen. Project 1 mainly focus on AI techniques, while Project 2 mainly focuses on energy storage, as well as EVs and its integration with power grid. In this project, we mainly focus on waterborne and railway transportation.

In waterborne transportation, we have investigated 1) a data-driven approach for ship trajectory length prediction in intelligent traffic signaling in controlled waterways. We have for the first time investigated the prediction of the overall trajectory length of manually controlled ships. We first group ships' historical trajectories by using the fuzzy c-means clustering algorithm, while the relationship between some known factors (i.e., ship speed, loading capacity, self-weight, maximum power, ship length, ship width, ship type, and water level) and the resultant memberships are then modeled using artificial neural networks. The trajectory length is then estimated by the sum of the predicted probabilities multiplied by the trajectory cluster centers' length. The experimental results show that the proposed method can reduce the probability of generating incorrect traffic control signals by 74.68% over existing signaling systems, and hence can significantly improve the efficiency of traffic management systems and increase the traffic capacity by reducing the traveling time [75]. 2) We have further investigated long-term speed prediction of ships in busy inland waterways, considering that inaccurate ship speed prediction leads to poor traffic signaling, hence causes significant traffic jams. We have used novel artificial neural network models for the ship speed prediction, achieving more than 97% accuracy in real-life experiments [76].

In railway, our focus is on the power supply of traction systems, in particular power electronics control to eliminate negative sequence and harmonics, to improve power quality and to integrate renewable generations. 1) We have investigated a modular multilevel converter (MMC) based static frequency converter station with renewable energy access. Wind power generation is coupled into the station via dc link of the back to back converter. The dynamic single phase traction load and intermittent renewable generation bring double frequency oscillation and large deviation problems to the dc link voltage. Special design considerations and control schemes are proposed for the MMC to stabilize dc link voltage by controlling the total number of total inserted modules, and to resolve the voltage oscillation issue caused by single phase load and stabilize dc link voltage under 10MW step change [77]. 2) We have investigated back-to-back converter-based railway power supply system (RPSS) for elimination of neutral sections on the traction side and improve power quality on the grid side, while accepting significant penetration of renewable energy and considering power mismatches. We proposed a novel power electronics topology which breaks the limit of back-to-back structure and enables more flexible free energy flow [78]. 3) We investigated a new co-phase system with a hybrid power quality conditioner (HPQC) for solving power quality issues caused by single-phase traction loads, while the renewable energy is proposed to integrate with the railway power supply, leading to a more complex system to model, design and control. We formulated the design of the compensation scheme in the cophase system as a multi-objective optimization problem which is solved by the NSGA-II algorithm to achieve optimal compensation. To further eliminate the effect of real-time error arising from load and renewable generation predictions, a hybrid optimal compensation control is proposed, leading to full and optimal compensations. Our studies show that new scheme is able to achieve an average of more than 20% reduction of the HPQC operation capacity, meanwhile the power quality is always satisfied even if in the presence of real-time prediction errors [79]. 4) We further investigate the railway co-phase traction power supply system (TPSS) with a power flow controller (PFC) to address the power quality and neutral section issues. To collect the regenerative energy and achieve a more flexible power flow, the energy storage system (ESS) is integrated into the co-phase system. As the key components, the reliability of power electronics modules in PFC and battery cells in ESS is highly related to their thermal performance. We have therefore investigated the operational thermal dynamics of power electronics modules and batteries, and developed a novel thermal constrained energy management strategy for co-phase systems integrated with batteries, reducing the peak traction power supply by up to 42.0%, and achieving up to 94.1% thermal reduction [80].

Our most recent interest in this area is the net-zero transformation of railway stations and feed stations using novel power electronics solutions and microgrid solution [81], and holistic approaches in planning and operation of transportation systems coupling with low-carbon power networks [82].