Development of a hybrid electricity-loss forecasting model based on DBSCAN clustering and MLP neural networks

Abstract

Relevance: operational data from distribution networks contain noise, outliers and locally heterogeneous operating regimes that may distort neural-network training and reduce forecast robustness. This creates a need for an integrated architecture in which density-based structural filtering and local nonlinear forecasting are treated as complementary stages of one predictive process.

Aim: to develop and validate a hybrid electricity-loss forecasting architecture that combines density-based DBSCAN clustering with local multilayer perceptron models, and to quantify the accuracy gain attributable to the preliminary structural decomposition of the data.

Methods: real operational data were subjected to density-based structural filtering using DBSCAN. The input feature space comprised active power, ambient temperature and relative humidity. PCA-assisted analysis was used to select ε=0.4 and minPts=5. A separate MLP was trained for each detected dense cluster, and the final forecast was generated through a composite Y_hybrid structure. Forecast accuracy was evaluated using MAPE.

Results: the baseline MLP achieved a MAPE of 2.58 %, whereas the DBSCAN–MLP architecture reduced MAPE to 2.10 %, corresponding to an 18.6 % relative error reduction. Actual-versus-predicted trajectories remained closely aligned for training, validation and test subsets, supporting the robustness of the hybrid architecture.