Posters

Abstract

The increasing demand for energy and climatic changes shifted global focus towards renewable energy. Wind energy is one of the best resources in terms of capacity factor and annual energy generation. Accurate wind data measurement is paramount for reliable wind power prediction. Wind data collected through sensors mounted on fixed buoy masts, forms the foundation of wind power forecasting models. Commercially operated wind farms have high hub-height wind turbines i.e., 100 m and more, and conventional statistical methods such as Power law and Logarithmic law for measurement of vertical wind speed are not precise solutions to this complex issue. The reasons for high wind speed at height are due to less influence of human activities, less surface roughness, stable atmospheric conditions and less friction losses. Accurate wind profile information is required at turbine's hub height for the estimation of energy which is generated from installed wind turbine. Several researchers have proposed different models for vertical wind speed extrapolation at higher height including physical models, statistical approaches, numerical or machine learning techniques. Numerous studies have shown the potential that Machine Learning Algorithms (MLAs) are extensive scientific approaches to solving complex non-linear problems. A novel approach for vertical Wind Speed extrapolation, known as Particle Swarm Optimization Long Short-Term Memory (PSO-LSTM) neural networks is used in this study. The standard LSTM, in contrast, represents an advancement in the realm of Recurrent Neural Networks (RNNs). It incorporates a recursive network delay, facilitating the mapping of network output for the nth sample with respect to previous inputs. The PSO-LSTM model is trained in two stages. Initially, Levenberg-Marquardt (LM) method is used to train LSTM. This approach is based on its rapid convergence and reliability. In second stage, in addition to the LM algorithm, the Particle Swarm Optimization (PSO) technique is incorporated to identify the optimal parameter set that minimizes the cost function. The Feedforward neural network (FFNN) is used as a benchmark for the comparison of the performance of PSO-LSTM. The performance of the PSO-LSTM, FFNN, and LogLaw algorithms in terms of RMSE is compared. The observed trend in Root Mean Square Error (RMSE) values for the PSO-LSTM model demonstrates a gradual transition, ranging from 0.21 to 1.03 m/s, considering height interval spanning from 50 meters to 120 meters. The RMSE values for the Feedforward Neural Network (FFNN) exhibit a similar progressive alteration, extending from 0.28 to 1.12 m/s, across the same height range. Similar trends are discerned when examining Mean Absolute percentage Error (MAPE), including PSO-LSTM, LSTM, FFNN, and LogLaw, there is an incremental rise in MAPE values from 2.69% to 9.53%, 3.03% to 10.01%, 3.92% to 10.36%, 4.95% to 23.43% respectively. The PSO-LSTM method achieved the highest levels of accuracy, with R2values reaching remarkable percentages of 96.25%, 88.16%, and 78.21% at altitude levels of 60 meters, 90 meters, and 120 meters, respectively. These impressive results underscore the PSO-LSTM's efficacy in accurately estimating wind speed values and highlight its potential as a leading method for wind speed extrapolation.