Ensemble echo state networks trained collectively for reliable energy forecasting in facilities

Accurate and reliable short-term energy consumption forecasting remains a central challenge in modern energy systems, where heterogeneous patterns make uncertainty quantification particularly difficult. This problem is often addressed using ensembles of lightweight recurrent neural networks, such as Echo State Networks (ESNs), which can be efficiently trained individually and then averaged. We propose a novel one-shot ensemble technique for ESNs that constructs multiple diverse reservoirs while jointly training a single shared readout layer. This approach combines the diversity required for effective ensembling with the efficiency of collective optimization, enhancing two key ingredients for accurate and reliable forecasting. Comprehensive experiments demonstrate that our method consistently outperforms conventional ensemble techniques. On educational buildings, the proposed approach achieves an 87.3% improvement in accuracy and an 18.8% enhancement in uncertainty quantification. In manufacturing facilities, where system dynamics are more complex, our method achieves a 13.6% gain in accuracy and a 7.5% improvement in uncertainty estimation. Notably, it provides the best balance between precision and reliability compared to several state-of-the-art models. • Balance between accurate and reliable energy consumption forecasting in facilities. • Novel one-shot ensemble of echo state networks trains reservoirs collectively. • Comprehensive evaluation of forecast performance and uncertainty quality. • Optimal forecasting equilibrium for manufacturing factories patterns.