An Overview of relevant control methods employed in the field of energy/load control in hybrid electric vehicles. Index 1. Introduction 2.

HEV and HAULE UAVs similarities 3. Control strategies a. PID control b. Adaptive control methods c. Model Predictive Control d. Fuzzy Logic e. Rule based strategy f. Neural networks 4. Proposal a. Requirements 5. Summary 1. Introduction: With the increased possibility of employing a road vehicle which is not relying on an internal combustion engine as its only power source, a new need has arisen to control the power flow of various interacting sources, together providing a single output. Ideally each individual source should be used at its most efficient operating point, optimizing the whole vehicle operation. This makes the control of the individual components essential and requires attention to control criteria on a subsystem and global system level. In the context of hybrid electric vehicles (HEV), common components needed to be controlled are: Battery Ultra-capacitor IC engine Electric motor Fuel Cell General difficulties with HEV control include the need to know the future and current states of the system, such as the vehicle speed and required load. Using a typical drive cycle satisfies this need, however when operating in real conditions, optimization is more complicated and employing predictive methods is useful. Some of these predictive methods include route recognition/ road elevation and wind detection. Transferring this to a long endurance unmanned aerial system highlights the same requirements as for the HEV analysis. The knowledge of the future load cycle is useful to optimize the energy management of the power sources. Since a future sudden power burst will require a high state of charge of a capacitor, which has to be including in the control procedure.

2. Hybrid electric vehicle (HEV) and HAULE unmanned aerial vehicle similarities A HEV will require a combination of power sources which together can satisfy the performance requirements, such as endurance and peak power (i.e. vehicle acceleration and top speed). Batteries and fuel cells generally function as base power sources providing the electric motor drive unit. Batteries may function as energy and power sources, however their power density is too low to provide the potentially required burst power of sudden acceleration. The same holds true for fuel cells, which have a very slow response time and low peak power operational abilities. This makes it necessary to include capacitors, which are very power dense but hold relatively little energy. Alternatively an IC engine may be used as a peak power source when required, powering the drive train directly or driving an electric generator. For a high altitude ultra long endurance UAV, performance requirements may be similar, with a stronger focus on endurance. A long output of relatively low power is required when cruising or loitering, while during take off a period of high power is present. Additionally uncertainties, such as sudden manoeuvres and wind gusts, will require the possibility to provide peak power during the cruising period. Furthermore the payload on board will have power requirements as well which may be of variable nature and has to be satisfied. Again, due to the multiple sources employed, the control will be essential to improve performance. Possible configurations for HAULE UAVs include fuel cell-battery hybrids, as well as pure battery designs, which have shown some success. 3. Control Strategies Generally one can distinguish between rule based control methods and optimal control. Accuracy, complexity and reliability will make certain approaches more suitable to a given application. Furthermore the difference between reactive methods, which reacted to changes in current conditions and predictive methods, has to be made. A short description of a number of approaches will follow. Most of them have been applied in the field of HEV, making them potential suitable for UAV control with some adaptations. a. PID control This is a generic feedback control operator, which uses proportional, integral and derivative action on the error system, and modifies the control input through the feedback, accordingly.

U(t) is the output signal, e(t) the error (input-output) and Kp, Ki, Kd are the gains from the PID controller. These gains can be set to control the output, to reach a certain value. Overshoot, rise time and settle time, are parameters which need to be considered, as they define the competence of the controller for its application. The tuning of the 3 gains can be done manually, by looking at the response of the system to general input and, then tuning the gains accordingly. Additionally, methods such as the Ziegler-Nichols method allow following a process when selecting the gain values. PID control is generally limited to linear systems and has some other limitations with respect to noise affecting the derivative variation and the size of the integral term, non-recoverable overshoot. b. Adaptive Control methods This approach is concerned with a controller, which adapts its parameters according to variations in the operating conditions. It is different from robust control, (which guaranties stability for constant gains if the changes are within a certain bound), as it is concerned with changes to the gains itself. Convergence and robustness are therefore of importance and have to be considered in the analysis. Applications may be, self tuning of linear controllers, adaptive control of linear controllers for nonlinear processes and multi input output systems. c. Model Predictive Control This method has been widely applied in the field of process engineering (chemical plants, etc) in the 70s, 80s, and is now regaining popularity in a wide variety of new fields, including HEV control. This method is an online method, meaning that it will optimize while in operation. This fact used to pose a limitation for process with fast control responses, since computing power was not available. Nowadays control rates of 50 Hz can easily be achieved for systems with large number of variables. The method MPC is a form of control in which the current control action is obtained by solving, at each sampling instant, a finite predictive horizon open-loop optimal control problem, using the current state of the plant as the initial state. The optimization yields an optimal control sequence and the first control in this sequence is applied to the plant. This process is then

repeated, by shifting the predictive horizon by one step and using the new initial state for the calculation.

Here one desires to achieve a reference output (red) through a series of optimal control inputs. One predicts the output trajectory and optimal input series, and chooses the first input as the actual control action at this time step. At each step the cost function has to be recalculated

xi = controlled variable ri = reference variable ui = input variable w = weighting coefficients N = finite horizon steps.

d. Rule based/ Heuristic Based on data on power requirements, IC engine performance, emission and fuel burn maps, one may decide to develop rules to control the different components. This may be in the form of simple on/off control, base line approaches and time dependant variations. Heavy tuning of control parameters is required, as well as static data on the fuel burn of the ICE. Common constraints that need to be included are the SOC of the storage devices (Battery, Capacitor).

e. Fuzzy Logic This approach relates a input set to an output set through a number of linguistic rule based IF .. THEN .. expressions. The use of fuzzy variables eliminates strict boundaries and makes use of membership functions to allocate inputs to fuzzy sets.

- fuzzy variable for the degree of truth SOC The definition of membership functions is important and relates the degree of truth of an input to a given field. For example if an input is SOC = 0.2, it will have belong to the LOW and XLOW field with different truth value for each. (i.e. greater XLOW than LOW). The definition of these functions has to be based on the system itself and the values for extreme operation conditions. A large number of rules may be required, which can be represented in rule matrices:

Here a positive large load is required, and the output of the fuel cell is dependent on the SOC of the battery pack (y axis) and the capacitor (xaxis). Low SOC for both battery and capacitor will result in a max power output of the Fuel Cell etc.

After producing the fuzzy outputs in the membership functions, according to the rule matrix, the result needs to undergo Defuzzification. This may be achieved through a number of ways, where one simple approach is to map the total area of the fuzzy variables and take its centroid as the crisp output value.

f. Neural Networks Neural Networks can be used to approximate a function with multiple variables and of high complexity. The parameters which define the network can be adapted in a certain way to produce a desired output. This makes it possible to control a system through the adaptation or training of a neural network.

(1) An architecture using a single neuron has the following form:

Both w and b are adjustable parameters and p are the inputs with a, the outputs.

Any number of these may be combined in a layer and form the following system, which can be re-written in matrix form. W now is the weighting matrix of the layer with its size depending on the number of inputs and neurons.

(2) Multiple layers can then be combined to form a Network of functions. Each layer output is the input to a next layer, creating a network with an input layer, any number of intermediate (hidden) layers and an output layer:

Neural Network training through back propagation: During the training process, an example set of inputs and outputs are given, to adjust the weights W, and bias b accordingly. In the back-propagation method, this is achieved by computing the errors between the training output and the actual output and redistribute the weight values, in order to minimize this. For a simple single layer network, this error is easily computed and can be related back to the inputs through gradient methods. However when dealing with multi layer networks, the output error needs to be weighted back to the hidden layers, in order to produce an adjusted of these weights as well.

4. Proposal for HAULE UAV systems energy control Flow of power output between available storage sources and producing source has to be controlled effectively to ensure long operational endurance. This may be achieved through the use of a predefined rules strategy, which has some knowledge of common operational situations and their power requirements. The controller will then adjust the different source flow accordingly and deliver the required power with optimal operating conditions of all elements involved. a. Requirements Produce the required total power output when demanded by the loads at current flight situation. Quick response times of the controller to adapt effectively.

Sources may be batteries, ultra capacitors or fuel cells. They will be subjected to some form of operational constraints C and efficiencies Eff. Some Constraints:  ) ( 

States:

(Energy in sources) Inputs:

) ( (Power demand on source determined by load demand) ) )) Maximize ( (  

It is possible to implement full controller training procedure from a load profile dataset, or successive online training through learning algorithms. An initial set has to be established to have sufficient optimality at the beginning of the mission.