Extreme Learning Machine (ELM)

In ELM, the weights and biases are selected in a random manner, and using predictors leads to the best solution. In ELM, the architecture is determined by an automatic mechanism in which a mathematical model is used to attributes of data using randomly allotted weights and biases. ELM uses feed-forward and backpropagation ANN, which improves the rate of convergence, generalization, tuning, and best fitting of data. In Ref. [12], ELM is employed to estimate worldwide sun radiation. In the testing phase, estimations made by ELM were compared with multiple linear regressions (MLR) and other models. Using ELM, the mean error and root mean square errors were found to be lower than that of MLR and other methods. ELM finds its application in power converters based PV systems where the objective is to improve the frequency stability of grid. In Ref. [13], a virtual inertia-based ML method is proposed for grid integration of PV system with improved frequency stability. Reinforcement learning (RL) is one of the popular forms of ML. Its background lies in learning mechanism of alive entities. RL performs actions on its environment and manipulates for the maximization of the received reward. The controller with ML technique holds advantages over the PI controller, such as ability to self-learn, capability to resolve the stability of frequency, quicker dynamic response, and adaptive control. Simulation studies showed that an ML-based controller was able to reduce error in steady state by 27% while maximum deviation of frequency with respect to nominal frequency was limited by 0.1 Hz. In Ref. [14], an algorithm that was a combination of RL with sleep schedule for coverage is applied. The proposed algorithm works in two stages. In the first stage, a precedence operator is assigned in a group and the formation of nodes is completed. In the next stage, a learning algorithm is generalized to a multi-sensor learning group and nodes are directed to work in a collaborative manner while adapting to the changing environment. The algorithm completes the learning of the entire team by changing the role of the active nodes while placing others in the sleep mode. The experimental results of the proposed algorithm on wireless sensors-based PV system suggests that the balance of energy consumption is maintained among the nodes, which results in increased life of the network along with desired coverage.

The output power of PV systems is generally dependent on the weather and climatic conditions. In Ref. [15], SVM learning with satellite technology was employed for predicting the output power of PV system. This prediction model was able to predict: (a) availability and quantity of clouds and (b) irradiance. Satellite pictures taken over the past four years were fed to configure the output and input data sets for SVM learning. The proposed model showed not only excellent accuracy in prediction as compared to conventional ANN models but also that prediction data can be used in grid-connected operation as well as in management of energy in the grid.

Identification of the type of fault and its location by using conventional methods is sometimes difficult in PV systems. This is due to variation in atmospheric conditions

i.e. irradiance, shading, temperature, etc. In Ref. [16], a semi-supervised learning scheme was introduced for the detection, location finding, and classification of the fault, along with corrective measures taken to resolve the fault. The results exhibited that identification and correction of all the learned and unlearned faults were accomplished by the proposed method when PV arrays with prior experience were considered. The variations in output voltage were also minimized in the fault condition.

Nowadays, there is a rise in the demand for clean energy. The charging of electric vehicles (EV) is done using charging stations which are fed by PV systems. There is a requirement for the management of energy that can optimize the cost of operation and performance of energy storage systems. In this context, a new deep RL method that can handle time-varying data such as the charging status of the battery and data related to vehicle is proposed in Ref. [17]. This method provides the following advantages:

  • • Computation of the solutions of scheduling in EV charging stations
  • • Handling of time-varying data
  • • Ability to achieve desired performance
  • • Reduction in operating cost of charging stations

In Ref. [18], an ML-based electroluminescence (EL) imaging technique is proposed for PV systems. Using features of EL, prediction models of power and resistance were built for a PV system. The key benefit of these models is the requirement of EL imaging with PV module characterization and V-I curve only for the fast estimation of degraded PV modules. This technique offers other advantages such as more accurate prediction of efficiency, fast response, capability to electrical properties, and management and operation of PV systems. The developed models were found to be accurate in estimating the change in a series resistor, the degradation in the performance of a PV module, and the power of modules of different brands.

< Prev   CONTENTS   Source   Next >