Explainable deeply-fused nets electricity demand prediction model: Factoring climate predictors for accuracy and deeper insights with probabilistic confidence interval and point-based forecasts

Electricity consumption has stochastic variabilities driven by the energy market volatility. The capability to predict electricity demand that captures stochastic variances and uncertainties is significantly important in the planning, operation and regulation of national electricity markets. This study has proposed an explainable consumptiondeeply-fused nets electricity demand prediction model that factors in the climate-based predictors for enhancedaccuracy and energy market insight analysis, generating point-based and confidence interval predictionsof daily electricity demand. The proposed hybrid approach is built using Deeply Fused Nets (FNET) thatcomprises of Convolutional Neural Network (CNN) and Bidirectional Long-Short Term Memory (BILSTM)Network with residual connection. The study then contributes to a new deep fusion model that integratesintermediate representations of the base networks (fused output being the input of the remaining part of eachbase network) to perform these combinations deeply over several intermediate representations to enhance thedemand predictions. The results are evaluated with statistical metrics and graphical representations of predictedand observed electricity demand, benchmarked with standalone modelsi.e., BILSTM, LSTMCNN, deep neuralnetwork, multi-layer perceptron, multivariate adaptive regression spline, kernel ridge regression and Gaussianprocess of regression. The end part of the proposed FNET model applies residual bootstrapping where finalresiduals are computed from predicted and observed demand to generate the 95% prediction intervals, analysedusing probabilistic metrics to quantify the uncertainty associated with FNETS objective model. To enhance theFNET model’s transparency, the SHapley Additive explanation (SHAP) method has been applied to elucidate therelationships between electricity demand and climate-based predictor variables. The suggested model analysisreveals that the preceding hour’s electricity demand and evapotranspiration were the most influential factorsthat positively impacting current electricity demand. These findings underscore the FNET model’s capacity toyield accurate and insightful predictions, advocating its utility in predicting electricity demand and analysisof energy markets for decision-making.