Which of the following are characteristics of a decentralized control system?

4.3 Performance comparison

The decentralized control scheme has been tested to ± 10% disturbances in feed flowrate, temperature and composition as well as a simultaneous set point change in all product purities from 98 to 98.5% (mole frac.). Loop interaction is clearly demonstrated with the set point change experiment (Fig. 2 left). MPC performance is compared to decentralized control by carrying out the same simultaneous set point change.

The results are shown in Fig. 2 and clearly demonstrate the advantage of MPC in eliminating loop interaction and oscillations. Settling times for distillate and bottoms purity are significantly lower. For sidestream purity and prefractionator vapour composition the decentralized control scheme presents lower settling times.

5.7 Control types for MEGs

Numerous control methodologies for MEG have been proposed and studied. The centralized, decentralized, and multilevel hierarchical control of MEG have been discovered in previous theoretical and laboratory experimental research [29–34]. An illustration of these different control structures will be discussed in this section.

5.7.1 Decentralized Control for MEGs

Decentralized control methodology for the MEG can be summarized as follows, the individual energy sources has the right to share the demand as per their specific capacity and local control characteristics, see Fig. 12. Those are fixed during installation and planning phases. Consequently, it is difficult to make any re-scheduling for instantaneous energy production, or for each source to achieve optimum generation cost and emission conditions. This fact led to the underutilization of the energy sources, although they may have high efficiency and lower operating rates [35].

Fig. 12. Decentralized MEG control methodology.

5.7.2 Centralized Control for MEGs

The centralized control methodology for the MEG mainly consists of a central control system for the remote control of all energy sources in the MEG boundary, see Fig. 13. Optimal performance can be achieved by using a centralized control system, but it has a significant disadvantage on the reliability of an energy system where the central controller fails, most likely the overall energy system will collapse. The centralized control methodology relies on a communication network, where the speed and reliability of the communication system has direct impact on the MEG performance, reliability, and resiliency.

Fig. 13. Centralized MEG control methodology.

5.7.3 Multilevel Hierarchical Control for MEGs

Multilevel hierarchical control of the MEG provides a better methodology to overcome most of the obstacles that accompanying the centralized and decentralized control methodologies [36]. This control type has a significant role in achieving the optimum operation of the MEG system that is similar to a centralized control methodology but under lower speed and the reliability level of the communication network requirements, see Fig. 14. However, the main challenge of the hierarchical control methodology is the necessity for clear boundaries of control range and domain-based control levels [37]. In the hierarchical control the supervisory and predictive control levels generally depend on a communication network to achieve the MEG system optimization operation that is the same as a centralized methodology. But hierarchical has the advantage of a decentralized methodology, where the reactive control level is not dependent on the communication network. This feature immunizes the MEG from the loss of operation once a failure occurs in a higher-level control and/or network. Whilst the hierarchical may lose optimal performance during such hazardous event.

Fig. 14. Hierarchical MEG control methodology.

The MEG performs dynamic control over energy sources, enabling autonomous and automatic self-healing operations. During normal or peak usage, or at times of the capital energy grid failure, an MEG can operate independently of the capital grid and isolate its generation nodes and energy loads from disturbance without affecting the capital grid's integrity. A multilevel hierarchical control is proposed to provide autonomous self-healing supervisory control for the MEG system. The control architecture consists of three levels working together to achieve the overall operational goal.

Hierarchical control architecture

Hierarchical control architecture was realized in order to operate energy sources proficiently as well as to employ the MEG components efficiently. It embraces three levels, comprises an autonomous decision making level, a predictive control level, and a reactive control level. Each level has its own local objectives and they work together to realize a resilient operational performance. The higher-level controller involves a fault tolerant control formulation in order to deal with the uncertainty of hazardous conditions and to determine the best action for each subsystem. The predictive control level utilizes prescheduled operational timing to handle the district cooling units (DC) operation. The predictive control aims to shift the operation of the DC units to off-demand periods by charging the TES for further uses at on-demand peak periods. The lower-level controller is a load following control for the demand, which needs fast response. The quality of any control depends on the assumed information and control structures [36]. A central decision maker defines values of control based on available information collected from all subsystems. Nevertheless a centralized method might be difficult to realize in a large-scale system, where a process of transmission and transformation of information are more complicated. The decentralization of information and control structures is a reasonable solution to overcome this problem. Control problems with decentralized measurement information are the main element for hierarchical control. The decomposition of a large system into subsystems is mainly aimed at minimizing the required computations to further to reduce the amount of information required for a decision making level [38].

Previously the MEG was classified either as an islanded or grid-connected approach. Whereas, it is essential for a resilient energy system to modify an elastic MEG configuration capable of operating in both grid-connected and islanded modes [39]. This system opens the door on great challenges where establishing such systems requires an integration of different technologies, energy sources, energy storage, and energy management systems. In addition to safety issues, such as fault monitoring, predictive maintenance, or protection, which are fundamental principles for a MEG with a high level of self-healing capability.

This section's emphasis is on hierarchical control design with the employment of an adaptive neuro-fuzzy system for energy balancing and optimization purpose and to improve the resiliency of the MEG system. Hierarchical control design is also aimed at providing real-time backup methodologies in case of a hazard event in cooling generation and/or cooling load profiles in order to minimize its negative impact on the electrical energy system.

Design of adaptive neuro-fuzzy decision making method

Fig. 15 shows an adaptive-network-based fuzzy inference system (ANFIS) that has an optimized structure of 5 layers organized as follow 2:10:25:25:1.

Fig. 15. Optimized ANFIS architecture.

This structure was created from initial data using a MATLAB/anfiseditor environment. The Takagi-Sugeno-Kang fuzzy model-based ANFIS has been used with architecture of two inputs and one output, which is tuned online using a combination of least-squares estimation and back-propagation methods. The error between reference chillers operation and actual chillers operation is used to tune the neuro-fuzzy model parameters. The functions of each layer in the ANFIS architecture are formalized as follows [40,41]:

Layer 1: It is a fuzzification layer where each node is symbolized by a membership. Five Gaussian curve membership functions are designated to each input as shown in Fig. 16, and its node equations is given as follows:

Fig. 16. Gaussian curve fuzzy membership.

(19)Gaussianξcσ=e−12ξ−cσ2

where c is MFs center and σ is MFs width.

Layer 2: Each node in this layer is a multiplier that multiplies the input signals and forwards the result to the 3rd layer

(20)μi,j= μAiξ1.μBjξ2, i=j=1,2,3,4,5

This equation characterizes the firing strength of a rule.

Layer 3: Each node in this layer calculates the normalized firing strength of each rule as given in the following:

(21)μ¯i,j=μi,j∑i=15∑j=15 μi,j

Layer 4: In this layer each node is multiplied by tuned variable weights (a0i,j, a1i,j) as shown in the following equation:

(22) Oi,j=μ¯i,j⋅fi,j =μ¯i,ja0i,j+a1 i,j⋅ξ,i,j=1,2,3,4,5

Layer 5: Is the final output layer of fuzzy system. Output of the system is the summation of all incoming signals from layer 3, computed as follows:

(23)Y=∑i=15 ∑j=15Oi,j,i,j=1,2,3,4,5

Energy system procedure

Fig. 17 summarizes the procedure for energy production and flow for electrical, cooling and heating energy systems within a resilient MEG.

Fig. 17. MEG energy production flow chart.

3.2 Decentralized method

In the decentralized control strategy, each distributed energy source (generation or capacity) works freely utilizing measured local signals. Subsequently, no source is the reference and all sources are at a solitary control level. So removing or including a source makes no aggravation in other source operations [31–33]. This technique requires all resources to be dispatchable. Nondispatchable resources can be changed to dispatchable ones by including an energy storage component. There is no requirement for the correspondence system and the absence of a reference source builds the reliability of the system. These techniques manage the output voltage and frequency based on the active and reactive powers delivered by the inverter [34,35]. Subsequently, better reliability and flexibility can be accomplished in the physical area of DG units. Here, the agent-based decentralized control method, sliding-mode voltage control and droop controls, which are more common for decentralized methods.

3.2 Decentralized control

In the decentralized control structure, all subsystems of the plant are separately controlled by their own LCs. There may be some interactions between the subsystems so that the less interactions may lead to be closer to the suboptimal condition (Fig. 2.3B). In decentralized scheme, the role of all controllable agents and their LCs are the same, and none of them has more important role. Compared with the centralized control, this scheme has higher reliability because it has not a centralized controller with excessive communication. On the other hand, the decentralized scheme provides highest independence for controllable agent so that each agent has its own LC, which operates based on its local measurements' feedbacks. Thus, there is no need for fast two-way high bandwidth communication system and centralized controller, which were vital in the previous control scheme. This means that failure in each of the controllable agents or LCs will not result in collapse of whole plant. When a failure occurs, without redesigning of controller, one (or more) controllable agent(s) and the related LC(s) can be added to or removed from the plant [15,43,44]. Successful usage of communication technologies such as Wi-Fi and Zigbee [45] and effective algorithms for information exchange such as Peer-To-Peer, Gossip, and Consensus [46–49] result in practical application of decentralized control and management. Therefore, it is possible to implement frequency and voltage control, DERs power sharing, and their energy management in a decentralized approach. As shown in Fig. 2.3, the decentralization level may vary from centralized to fully decentralized [50]. Decentralized control can be realized using multiagent systems (MASs). They can improve the intelligence level of LCs by transferring decision-making capability to the elements in the local side. Here, the communication system plays an important role since the local decision-making is based on data received from the environment of each element and its adjacent devices. The decision making is locally processed, and energy management system must perform information sharing/coordination. Therefore, the decentralized approach decreases the computational burden. Also, after loss of MGSC/EMS, the system can operate further, which results in higher reliability than the centralized one. Another privilege is that it can provide a suitable framework for plug-and-play functionality, which is important for the application of plug-in hybrid electric vehicles (PHEVS). This capability improves MG flexibility/expandability. However, synchronization among the MG elements and information security in its communication systems are important challenges in decentralized mode of operation. In the following MGs, a decentralized control system can be considered as a good solution [51–53].

The MGs, which have a large size, or their DGs, ESSs, and loads are widely dispersed. In this case, data acquisition can be difficult or costly.

In case of frequent reconfigurations in MG due to need for plug-and-play operation mode of devices such as PHEVs and RERs, a decentralized control system is used.

The MGs, whose DERs belong to different owners and should be considered as different players in MG. They have their own operation objectives and decision-making procedures.

The decentralized control method has been realized using MAS [30,54], sliding-mode control [55,56], and droop control [25,31,32,57–61]. The DERs, in these methods, have access to their local measured signals and operate independently. Therefore, their plug-and-play operation is possible with less problems in comparison with centralized methods. In the droop control method, the voltage and frequency of DGs are controlled considering their active and reactive power exchange with the MG without using any communication links [62–65]. Therefore, higher reliability and better flexibility can be achieved [17,32,40], and as a result, many researchers have suggested the application of the droop control method [66] and its variants. Also, almost all the experimentally implemented MGs are based on this method [67].

1 Introduction

Recent designs of cooperative systems involve decentralized and distributed controls that communicate locally rather than globally. Designs for physical systems with underlying communication networks and cooperative controls have been developed by many different techniques. Some techniques include the use of passivity as a design tool [1], Lyapunov methods involving passivity [2], and graph theory methods [3]. Systems with a shortage of passivity lie in a broad category of dissipative systems which include non-minimum-phase systems and systems with relative degree greater than 1, whereas passive systems include systems of low relative degree and minimum phase as shown in [4] and [10]. The utilization of passivity indices allows controls for switched systems to be designed as shown in [8]. Passive systems may be connected under certain conditions, such as implementation of negative feedback of two passive systems, or connection of passive systems in parallel. These properties are valuable for networking passive systems; however, if a system is not passive, the same properties of connection will not hold. Differences in passive systems and passivity-short systems and their respective properties will be investigated in terms of stability and input-output analysis in Sections 2 and 3. Some variations of nonpassive systems, specifically systems that have a shortage of passivity, or passivity-short systems, possess desirable connectivity properties that may be used in a networked design of such systems. A specific subclass of passivity-short systems—namely, input feedforward passive systems—are the most desirable type of passivity-short systems for networked operation because of their input-output relationships, as shown in Section 4. Methods for designing controls for input feedforward passive systems are the subject of Section 5. It is shown in Section 6 that there are various ways to interconnect passivity-short systems, and in Section 7 it is shown that by use of the characteristics described in Section 4 and combination of the design method in Section 5 with network-level controls, the overall system can achieve consensus, and by design, becomes a plug-and-play system.

1.2.3 Forecasting

The implementation of both centralized control and decentralized control strategies requires load forecasting, actual renewable resources power output estimation and information about electricity prices. Forecasting load and renewable generation requires data collection and a weather forecast, which may increase the operation cost [4,5]. Therefore, the benefit of forecasting should be greater than the extra cost involved. Currently, forecasting is mainly focused on electricity price, load demand and PV generation aspects for large interconnected systems. A number of forecasting methods are available and implemented in utility operation. The simplest forecasting method (persistent method) is to predict a variable based on its current value.

1 Introduction

Recent years have witnessed increasing interest in decentralized control of multiagent systems (MASs) [1–9]. Decentralized control is characterized by local interactions between neighbors. In comparison with centralized control, decentralized control avoids a single point of failure, which in turn increases robustness of MASs, allows use of inexpensive and simple agents, and lowers the implementation cost. In addition, decentralized control scales better as the number of agents increases and is sometimes an intrinsic property of MASs. The problem of synchronizing agents’ outputs is a typical problem solved in a decentralized fashion [2, 5]. The goal of output synchronization is to achieve a desired collective behavior of MASs. Examples of this are formation control, flocking, and consensus control.

Information exchange among neighbors is instrumental for coordination as discussed in all aforementioned references. However, degraded communication environments are commonly encountered in real-life applications. Namely, the exchanged information is often sampled, delayed, corrupted, and prone to communication channel dropouts. Furthermore, the agent models used to devise cooperative control laws might not be accurate or external disturbances might be present. Hence it is of interest to determine how robust a control law is with respect to realistic data exchange and to quantify this robustness. Herein the impact of external disturbances is quantified through Lp-gains, while the influence of corrupted data is quantified through the bias term in Lp-stability with bias. As expected, different transmission rates and communication delays as well as different controller parameters yield different robustness levels. In addition, since many MASs possess a set of equilibria (rather than a sole equilibrium point), we employ the notions of Lp-stability (with bias) and Lp-detectability with respect to a set [10, 11].

The salient features of this chapter are threefold. First, our analysis applies to MASs with general heterogeneous continuous-time linear agents (not merely to single- or double-integrator dynamics) with exogenous disturbances, directed communication topologies, and output feedback. The settings and goals of [5, 7] appear to be the most similar to ours. However, Yu and Antsaklis [5] consider passive agents without external disturbances, which in turn implies that the number of inputs equals the number of outputs for all agents, and balanced topologies, while Liu and Jia [7] do not deal with transmission rates, directed graphs, or output feedback. Unlike Yu and Antsaklis [5], we do not impose specific requirements on the agent and controller dynamics per se nor on the underlying communication topology. Basically, when given local controllers do not yield Lp-stability (with respect to a set) of the nominal system (defined precisely in Section 4.1), one can seek another topology or design alternative controllers. Second, when communication delays are greater than the sampling period, no other work takes into account the concurrent adverse effects of all aforementioned realistic data exchange phenomena [12]. For plant-controller settings, comprehensive theoretical and experimental studies of degraded communication environments are reported in [13, 14]. Nevertheless, these references impose zero-order hold (ZOH) sampling and do not consider noisy data or scheduling protocols [11, 12]. In addition, Kruszewski et al. [13] do not take into account disturbances. Third, the transmission intervals obtained are confirmed experimentally (see Fig. 1). It is worth mentioning that parts of this chapter were published in [15].

The remainder of the chapter is organized as follows. Section 2 provides the notation and terminology used herein. Section 3 states the robustness problem of linear MASs with respect to realistic communication. A method to solve this robustness problem is presented in Section 4 and experimentally verified in Section 5. Conclusions and future directions are given in Section 6. The proofs are provided in the Appendix.

8.6 Conclusions

In this chapter, we discussed the decentralized control of the large-scale system using a model reference adaptive control strategy considering constant state and input delays and time-varying delay associated with the interconnected term. We studied three methods for input delay compensation in a large-scale system. In the first method the interconnected term is bounded using a polynomial function with unknown coefficients. In the second method the interconnected term is bounded using a function with an unknown upper bound, and the third method examines an unstable open loop with a linear and unknown interconnected term.

In this study, we proposed two new solutions to compensate the input delay; first, a nested predictor is established to predict the future states for input delays compensation. Then by assuming that the predicted state is available the adaptive controller is designed. Also, the nonlinear interconnection terms are assumed to be bounded by polynomials with unknown gains. This is a step-by-step approach adopted to find a proper prediction state. The approach can be extended for using in the large-scale systems because, in such systems, the interconnected term is practically unknown, although it needs to be known for exact prediction. To solve this problem, we proposed an adaptive observer assuming that the interconnected term is bounded with a function that can be unknown. The purpose is to reach more accurate predictions and, finally, suitable tracing by knowing the interconnected term in the observed system. The next method uses dynamic compensation in which the input delay problem is converted into the state delay problem. Also, for systems that use model reference adaptive control, it proposes a simple and practical method. A delay can be used instead of prediction.

The features and limitations of each method with respect to simulation results can be summarized as follows. A nested predictor can present a suitable response to compensate input delay, but it may cause a delay in practical applications when solving problems numerically. The “adding integrator” method is a more appropriate solution than a nested predictor because it can be used in open-loop unstable systems and avoid solving system problems numerically, but it adds new poles to the system and diminishes the delay compensation speed. Nonetheless, each method has its own advantages, which depend on the system characteristics in practical applications. Interconnected term bounding is a characteristic index that separates such approaches and solutions, where the second method has superior performance because it can bound this term using a function with an unknown upper bound. Time-varying delays in nonlinear interconnected terms are bounded nonnegative continuous functions, and their derivatives are not necessarily less than one. Accordingly, we can say that the second method can cover more systems in practical applications. Adopting such choices gives the designer more authority in a practical system so that the operator can first determine for which items proposed in this research the system shows more adaptation and then gain more acceptable results by adopting an optimal strategy.

Some points of distinction between this and the previous studies are that the input delay and bounding interconnected term with the time-varying delay were considered simultaneously. Using the adaptive control strategy is another important aspect of the present work. Due to self-learning capability, this controller is a proper solution for many systems, which makes them robust. In cases where there is uncertainty in the system robust control is another choice depending on the system type. The reasons for choosing an adaptive controller for large-scale systems are the feasibility of tracking the reference model and estimating unknown upper bound in the interconnected term with less conservative conditions in comparison with other robust methods. On the other hand, robust control consistently deals with the worst state, and therefore it has larger dimensions and thereby more complexity for practical applications. The validity of the main results was verified through several numerical examples. Therefore the proposed methodologies apply to a class of interconnected large-scale systems with time delays.

Abstract

This chapter presents a control scheme for the decentralized control of power system dynamics. The scheme utilizes dynamic state estimation (DSE) using local measurements and machine parameters and employs the concept of pseudoinputs for decentralization. The method is based on the extended linear quadratic regulator (ELQR) and adapts in real-time to varying operating conditions of the system. The method is also computationally feasible and easily implementable. Using an example test system, it has been shown that the integrated scheme of DSE and ELQR can be utilized for dynamic estimation and control of small-signal dynamics of power systems in a decentralized manner, and it has several advantages over other linear control methods.