Clustering Hybrid Application for Load Forecasting in Smart Grids

• Published in: Applications of Big Data and Artificial Intelligence in Smart Energy Systems Vol. 1; pp. 69 - 100
• Main Authors: Peñaloza, A. Apolo, Leborgne, R. Chouhy, Balbinot, A.
• Format: Book Chapter
• Language: English
• Published: River Publishers 2023
• Edition: 1
• Subjects: General Engineering & Project Administration; General References; Software Engineering; Weight Update; Short Term Load Forecasting; RNN; LPWAN; Autoencoder; DE Algorithm; Convolution Layer; Pool Layer; Learning Rate; Elbow Method; big data; K-means; Gaussian Mixture Model; LSTM; Load Forecasting; Moment Rate; GMM; Dataset Pre-processing; Neural Networks; Hybrid Renewable Energy System; Non-zero Center; Clustering Methodology; Hyperparameter Selection; Deep Neural Network; Neural Network Performance; Smart Energy System; Forecast Methodologies; clustering, GMM; PAM; Time Varying Electricity Pricing
• ISBN: 9788770228251; 9788770229944; 8770228256; 8770229945
• DOI: 10.1201/9781003440710-4

Load forecasting is a key element for the correct operation of power distribution networks. Inaccurate load forecasts cause network overload, power outage, and end-user discontent, among other negative impacts. Several factors add uncertainty to load forecasting. The load forecast on microgrids in island mode is particularly important for their consumers, as these consumers need to balance their load and generation availability. In the last decade, a great number of papers have focused on different methodologies for short-term load forecasting. Many researchers concluded that data-based feature extraction is necessary for the best performance of forecasting methodologies. Thus, this chapter aims to analyze different methodologies for feature extraction as a pre-processing stage. These methodologies are K-means clustering, partitioning around medoids (PAM), and Gaussian mixture model (GMM). An autoencoder with convolution neural network (CNN) architecture is also analyzed. These pre-processing techniques are evaluated when used by forecasting methodologies such as long short-term memory (LSTM). This chapter also presents a combination of hybrid methodologies such as: LSTM-K-means, LSTM-PAM, LSTM-GMM, and LSTM-autoencoder. The case study uses a public dataset of Ireland, which includes per-hour measurements for 18 months from 256 consumers. The results show that LSTM performs better with feature extraction than with hyperparameter selection. The best results are obtained by autoencoder methodologies that reduce the data dispersion and improve the forecast performance. Many researchers show that data-based feature extraction is necessary for the best performance of forecasting methodologies. Thus, this chapter aims to analyze different methodologies for feature extraction as a pre-processing stage. These methodologies are K-means clustering, partitioning around medoids (PAM), and Gaussian mixture model (GMM). An autoencoder with convolution neural network architecture is also analyzed. These pre-processing techniques are evaluated when used by forecasting methodologies such as long short-term memory (LSTM). This chapter presents a combination of hybrid methodologies such as: LSTM-K-means, LSTM-PAM, LSTM-GMM, and LSTM-autoencoder. The case study uses a public dataset of Ireland, which includes per-hour measurements for 18 months from 256 consumers. The results show that LSTM performs better with feature extraction than with hyperparameter selection. The best results are obtained by autoencoder methodologies that reduce the data dispersion and improve the forecast performance.