Role of Artificial Neural Networks in Microgrid
The increased prevalence of distributed generation resources, incentives and the need to provide a reliable, secure and clean supply of power to consumers has led to the emergence of a new type of power sub-system, the "Microgrid". Microgrids address concerns about stability and power quality, with self-sustainability and fault tolerance. They do, however, raise new issues and challenges regarding effective power system operation and control. Though conventional techniques can resolve most of these issues, Artificial Intelligence techniques usually provide more robust solutions. They also speed up development time.
This article presents our point of view on the need for Artificial Neural Networks (ANN) in Microgrids and describes the role of ANN in realizing some of the functions of a Microgrid.
The inter-connection of a large number of renewable energy sources to the grid causes problems of control and stability, and also raises concerns about reliability, power quality and security. To resolve these concerns, the preferred approach is to view Distributed Energy Resources (DER) and their associated loads as a single entity, the 'Microgrid'. Essentially, Microgrid is a smaller portion of the power system. It accomplishes enhanced reliability of supply, quality of power and optimal cost of operation using a combination of DER (like solar photovoltaic cells, small wind turbines and fuel cells), storage (like batteries and fly wheels) and loads.
In a conventional grid, utilities manage the balancing of demand with supply, as well as handling other functions like stability and power quality. In a Microgrid environment, however, these functions are largely self-managed 1, in one of the following ways:
- Decentralized control -- is based on the principle that decisions should be taken locally, by the controllers, as centralized control is difficult. Different agents like load agent, power seller agent, and DER agent, communicate with each other to perform the common goals of Microgrid operation.
- Hierarchical control -- is achieved by using three levels of controls -- System, Master and Slave (Micro source and Load controllers). The System controller provides information like market signals and emergency conditions to the Master controller. The Master controller performs various optimization functions for the reliable and economical operation of the Microgrid. Slave controllers perform local settings based on the commands received from the Master controller.
Microgrid operates in three modes: Normal, Emergency and Island. Some general functions like optimization, performance assessment, and protection, are common to all modes of operation. Specific functions, like Black start capability, are seen only in the Island mode of operation.
Artificial Intelligence plays a vital role in Microgrid operation, allowing it to provide better solutions to different problems. Though Artificial Intelligence techniques include Artificial Neural Networks, Fuzzy Logic (FL), Genetic Algorithms (GA), Multi Agent Systems (MAS), and Expert Systems (ES), this article focuses on the application of ANN to solve Microgrid problems.
Artificial Neural Networks
ANN are composed of simple, highly inter-connected processing units -- neurons. These work in unison to solve specific problems. They are non-linear by nature, with excellent features like fault-tolerance and the capability for self-learning, which makes them robust and perfectly suited for parallel processing. ANN are also good at extracting the relationship between input and output variables where well-defined or easily computable co-relations do not exist, provided there is enough data for training.
ANN techniques are relatively easy to implement and do not require any prior knowledge of the system model. They include powerful pattern recognition capabilities and are flexible to model, making them attractive for solving real world problems. While applying ANN techniques to any particular problem, there are issues with respect to data, modeling and training which need to be considered.
Need for Artificial Neural Networks in Microgrid
Microgrids are characterized by self-sustainability, fault tolerance, reliability, security and power quality. To achieve these objectives, efficient, fast, and scalable optimization and control algorithms are required. These algorithms should be capable of processing information intelligently and taking critical decisions dynamically.
Though conventional techniques are successful in solving most of the problems in Microgrid domain, there are situations where they lead to unsatisfactory results. These include:
- Various forecasting tasks, like renewable energy forecasting, storage forecasting and demand forecasting, that need intelligent rules.
- The use of new equipment like storage systems, where monitoring and mapping of faults to different fault conditions of the equipment is difficult.
- The use of new equipment like power electronic interfaces, where monitoring and mapping of faults to different fault conditions and development of control mechanisms is challenging.
- The renewable energy system planning stage of the Microgrid, where the algorithms developed need to incorporate the practical knowledge of power engineers.
- The inclusion of renewable energy sources, for which the calculation of generation units to be committed and the economic scheduling of these units for optimal operation is highly complex.
Hence the conventional way of modeling the algorithms for these types of situations needs to be augmented or replaced with intelligent techniques that are robust and fault-tolerant. ANN provides a promising alternative in the situations outlined above.
Application of Artificial Neural Networks to Microgrid Functions
Power system problems can be classified as non-linear, dynamic, discrete, stochastic and random. Of these, non-linear problems are difficult to solve and ANN techniques are well suited to provide better solutions 2.
The role of ANN in solving non-linear problems like Forecasting (Demand and Renewable Generation), Protection, Intelligent diagnosis and Stability of Microgrid is discussed below.
One of the primary objectives of the Microgrid is to operate economically and reliably. This depends greatly on accurate demand forecasting and renewable generation forecasting.
- Demand Forecasting -- The purpose of demand forecasts is to predict the power that will be consumed by the load. In conventional power systems, demand depends on the weather, type of day and random activities or incidents. There are various statistical and Artificial Intelligence techniques which can forecast the load for the next 24 hours with a high degree of accuracy (in the order of 1-5% error).
In the case of a Microgrid, the load profile is not smooth, and the uncertainty increases as the Microgrid gets smaller. Due to high variations in the load profile, it is difficult to accurately forecast demand. Hence, the models developed should try to predict the load with reasonable accuracy. Some of the factors to be considered while developing the ANN model are:
- Increased responsiveness of demand to predicted real-time electricity prices as compared to conventional grids.
- Classification of load profiles of different customer classes based on the consumption of electricity.
- Need to input past demand as different components (like base load, peak load, valley load, average load and random load), instead of combining them into a single input.
- Dependence of the type of Microgrid on specific input variables. e.g., Type of day (weekday/weekend) input may not have much effect on forecasting for a Hospital Microgrid, whereas it is a major factor impacting the load profile of a Residential Microgrid.
- Renewable Generation Forecasting -- In a Microgrid, most of the generation sources are renewable, where the output depends on meteorological parameters like sunshine, wind speed and temperature. The complex interaction of these parameters makes renewable generation forecasting a challenging task, prone to significant forecasting errors due to the highly stochastic nature of weather forecasts. Hence an accurate prediction model is required.
- Wind Power Forecasting -- In general, there are 2 types of wind power forecasting. The first one forecasts the wind speed and converts it to power using an empirical formula. The second one directly forecasts wind power generation. There are various physical and statistical models for wind power prediction of up to 3 days ahead. The physical models require mathematical modeling of various energy sources -- a complex task. The statistical models aim to find the relationship between meteorological predictions, historical measurements and output generated.
As wind speed is dynamic and intermittent, and a high degree of non-linearity exists between wind speed and the power generated, we suggest a hybrid neural network model with 2 modules to forecast wind power accurately. The first module predicts the wind speed and this can be converted to wind power using an empirical formula. The second module predicts the wind power directly. The outputs of the two modules can be combined recursively to arrive at an accurate final forecast.
- Solar Power Forecasting -- Solar power forecasting is dependent on multiple inputs like solar radiation, installation angle, temperature, atmospheric pressure, relative humidity, time, and day type. All of these do not have a significant impact on forecasting at all time intervals. In order to determine the effect of these multiple inputs on the forecasted solar power output, we propose an integrated architecture with different models, each of which has unique inputs. Each model forecasts a solar power output and all the outputs can be combined recursively to deliver an accurate final forecast.
The presence of Distributed Generation (DG) in a Microgrid environment influences the design of the protection system. Conventional distribution networks which have unidirectional current flow employ over-current protection schemes. But, with the inclusion of DG, network power flows may change, making conventional protection schemes unsuitable. Moreover, as the Microgrid acts as a single entity to the main grid, the protection scheme should include the capability to isolate itself in case of a fault in the main grid. Similarly, the protection scheme should be capable of localizing faults within the Microgrid. In particular, the sensitivity and selectivity of the protection system needs to be carefully designed to ensure secure and reliable operation of the Microgrid.
Conventional techniques involve complicated calculations and may introduce errors in the estimated fault distance 3. These can be overcome by the application of ANN. For the reasons mentioned below, we feel that ANN are suitable for designing an efficient Microgrid protection scheme.
- ANN consider the whole input space in protection systems, while conventional fault classification algorithms do not.
- ANN have emerged as a pattern recognition technique, perfect for solving pattern recognition problems like distance protection.
- ANN are, by nature, fault tolerant and capable of handling corrupt data, making them good at fault detection and fault location.
- ANN determine fault direction accurately which is vital due to the intermittent generation and periodic load variation inherent in Microgrids.
As the inputs of the neural network model may affect the operating boundaries of the relay characteristics, proper insights for understanding the impact of inputs on relay characteristics need to be given.
Intelligent Diagnosis of Equipment in Microgrid
All electrical equipment are subject to incipient faults. These have to be monitored before they actually occur, so as to avoid equipment failure, which results in substantial financial loss. The incipient faults can be detected using diagnostic techniques by observing the degradation in system performance. Due to the nature of observed data and available knowledge, diagnostic methods are often a combination of statistical inference and machine learning methods.
ANN are a better option for diagnosing faults in electrical equipment for the following reasons:
- They can interpolate from previous learnings and give a more accurate response to unseen data, making them better at handling uncertainty.
- They are fault tolerant, so they handle corrupt or missing data more effectively.
- They are good non-linear function approximators by nature, making them better at equipment diagnostics.
- They are more suitable for extracting the relationship between input and output in fault detection and diagnosis applications.
Stability of Microgrid
The time varying dynamics of an electric power system are non-linear in nature and modeling these dynamics with conventional techniques is challenging. Moreover, these techniques are not applicable to the entire spectrum of operating conditions. The integration of renewable energy sources adds another dimension of complexity to the design of various conventional controllers in a Microgrid. And the data obtained from various sensors can be faulty due to bad connections and communications or hardware failures.
As ANN are robust and fault tolerant, they are used to design efficient power system controllers -- for PQ control, droop control and frequency/voltage control -- that stabilize the Microgrid during disturbances. These controllers:
- are efficient at handling slow and fast variations, generation variation and fast islanding arising from the inclusion of renewable energy sources in Microgrids.
- are efficient at handling disturbances caused by various power electronic interfaces in Microgrids, through a faster controller mechanism.
- are effective at tuning controller parameters with the non-linear and stochastic equipment used in the Microgrid.
The concept of Microgrid is gaining ground and demanding new computational technologies for efficient, reliable and secure operation, and control. With the inclusion of intermittent renewable energy sources and dynamic loads, operation and control of the Microgrid becomes complex. Hence, conventional modeling techniques need to be augmented or replaced with Artificial Intelligence techniques to handle Microgrid operations securely, efficiently and economically.
In this article, we presented our point of view on the role of Artificial Neural Networks in realizing some of the functions of a Microgrid. We recommend further research into this topic for an in-depth understanding of suitability of various ANN algorithms.
References 1. M. A. Pedrasa, 'Overview of Microgrid Management and Control', Energy Systems Research Group, University of New South Wales. 2. Raj Aggarwal and Y. Song, 'Artificial Neural Networks in Power Systems -- Part 3', Power Engineering Journal, December 1998. 3. K. K. Li, L. L. Lai and A. K. David, 'Application of Artificial Neural Network in Fault Location Technique', International Conference on Electric Utility Deregulation and Restructuring and Power Technologies, London, April 4-7, 2000.
1. M. A. Pedrasa, 'Overview of Microgrid Management and Control', Energy Systems Research Group, University of New South Wales.
2. Raj Aggarwal and Y. Song, 'Artificial Neural Networks in Power Systems -- Part 3', Power Engineering Journal, December 1998.
3. K. K. Li, L. L. Lai and A. K. David, 'Application of Artificial Neural Network in Fault Location Technique', International Conference on Electric Utility Deregulation and Restructuring and Power Technologies, London, April 4-7, 2000.