Powernet: A novel method for wind power predictive analytics using Powernet deep learning model

Abstract

Sustainable energy is a significant power generation resource for a cleaner and CO2 free environment. Out of different renewable energies out there, wind energy is rapidly growing sector and integrated to power grid. However, uncertainty,stochastic and non stationary nature of meteorological features, on which wind power depends, makes it difficult to predict accurately. Efficiency of wind farms and the power grid is directly proportional to efficient wind power predictive analytics. This study describes a hybrid model named PowerNet for improving the predicted accuracy in the field of wind power analytics. The improved hybrid model is a combination of Convolution 1 Dimensional and Bidirectional Long short memory (BiLSTM) models. Firstly, Conv-1D layers extract the spatial features of timestamped data sequentially. Then the output generated by multiple convolution operations at the nested layers is embedded with BiLSTM to work on the temporal characteristics of wind power data. The nesting of spatial and temporal extractor generates a novelarchitecture, Powernet for wind power forecasting from raw data. The effectiveness of powernet has been validated on real time wind power NREL dataset. Also, error and computational analysis has been conducted for short-term wind power forecasting with an ensemble of LSTM based models. The comparative analysis demonstrates that the proposed model powernet achieves better prediction than traditional deep learning standalone and hybrid models. Also, the statistical models are compared to show the raw data needs to be pre-processed when conventional models are applied. However, Powernet does not need the overhead of pre-processing for generating better predictions.

Winds of Change: Enhancing Wind Power Generation Forecasting with LSTM Models and Advanced Techniques in a Single Turbine Wind Farm

“…CNN, known for its ability to learn from image data and make intelligent decisions, has been utilised for time series classification and hybridised with LSTM for time series forecasting. Attention-based models, such as Attention-based LSTM, GRU, and encoder-decoder, have been applied in several studies (Garg & Krishnamurthi, 2023). DBN has been used for WP forecasting and its performance has been quantified using metrics such as MAE, SDE, and RMSE.…”

Short-term Wind Power Prediction based on Combined LSTM

“…As an improved model of recurrent neural network, long short-term memory (LSTM) network replaces ordinary neuron modules with special memory neurons to enhance the ability to remember the long-term training state and effectively solve the problem of gradient disappearance and explosion in recurrent neural networks to alleviate the problem of falling into the local optimum caused by overfitting [20][21][22][23]. Its network selectively exchanges information through the structure of information and control gates.…”