An optimized deep learning network using evolutionary algorithms for prediction of energy of piezoelectric harvesting system

Abstract

Complex dynamic electromechanical systems are known as newlinepiezoelectric energy harvesters, or PEHs. Deriving an appropriate model that newlinecan accurately represent the behaviour of a system under various operating newlinesituations is so difficult. The modelling procedure is also made more difficult newlineby the interaction between the piezoelectric transducer and any kind of power newlineconditioning and storage. This thesis holds the difficult task of modelling the newlinetransient behaviour of a PEH with electrical interface circuits while an newlineexternal storage device is being charged. Our novel simulation-based newlineoptimization approach can help you acquire the optimal geometric and circuit newlinedesign parameters to boost the efficiency of energy harvesting. newlineThe optimization platform is built upon the produced semi newlinetheoretical model of the energy harvesting system. Optimization challenges newlineusing a Deep Learning and Evolutionary Algorithm (EA) combination results newlinein an Objective Function that is costly to analyse and eats up valuable space newlineand time. To develop a robust neural network (NN) model, a collection of newlinetraining data derived from the simulation model is employed. This paper newlineproposes an improved deep learning method for Piezo-Electric Energy newlinegeneration forecasting using Bayesian Long-Short Term Memory (BiLSTM) newlineoptimization and Gradient Boost Regression (XGBoost). After the XGBoost newlinemethod extracts the complex and nonlinear relationships between time steps newlineand sequences, an attention weight vector is generated for the BiLSTM newlinehidden layer output. This enables the weighting of key factors at various time newlinesteps to influence the input. newline