Browsing by Author "Alp, Miray"

Now showing 1 - 2 of 2

A Novel Clustering-based Forecast Framework: The Clusters with Competing Configurations Approach
(2023) Gökhan Demirkıran; Miray Alp; Alp, Miray; Demirkıran, Gökhan
Accurate aggregate (total) short-term load forecasting of Smart Homes (SHs) is essential in planning and management of power utilities. The baseline approach consists of simply designing and training predictors for the aggregated consumption data. Nevertheless better performance can be achieved by using a clustering-based forecasting strategy. In such strategy the SHs are grouped according to some metric and the forecast of each group's total consumption are summed to reach the forecast of aggregate consumption of all SHs. Although the idea is simple its implementation requires fine-detailed steps. This paper proposes a novel clustering-based aggregate-level forecast framework so called Clusters with Competing Configurations (CwCC) approach and then compares its performance to the baseline strategy namely Clusters with the Same Configurations (CwSC) approach. The Configurations in the name refers to the configurations of ARIMA Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) forecasting methods which the CwCC approach uses. We test the CwCC approach on Smart Grid Smart City Dataset. The results show that better performance can be achieved using the CwCC approach for each of the three forecast methods and LSTM outperforms other methods in each scenario.
Akıllı şebekeler için kısa vadede yük toplam yük tahmini için kümeleme tabanlı derin öğrenme stratejisi
(2022) Alp, Miray; Demirkıran, Gökhan
Short-term load forecasting is of great importance in determining whether a smart home (SH) will participate in a Demand Response (DR) event created by the utility operator. However, volatility and uncertainty involved in energy consumption of SHs poses difficulties for their forecasting. Additionally, given the large number of SHs in DR programs of future smart grid scenarios, designing a unique predictor for each SH and constantly updating it faces big-data related challenges and may not be an economically viable approach. To tackle these issues, we evaluate three different methodologies, which are ARIMA, MLP and LSTM and different configurations of models are used. The proposed methodologies are tested on Smart Grid Smart City dataset and the date is picked to ensure that the greatest number of customers are available, after that, using k-means clustering the houses are grouped into clusters according to their informative features obtained from an extensive statistical analysis. Instead of having one predictor, each cluster has a pool of predictors that compete for the cluster's overall consumption forecast. Then they're added together to reach the final overall consumption forecast. We use evaluation metrics such the RMSE and MAE, which are scale-independent and robust to values near zero, to properly measure predicting accuracy at the aggregated level.