Hao Zhe Tan0000-0002-4082-3322Yun Seng Lim0000-0002-7211-9524Jianhui Wong0000-0003-4889-4907Lee Cheun Hau2026-10-132026-10-132026-1010.1016/j.est.2026.123158https://dspace-cris.utar.edu.my/handle/123456789/11713Maximum demand (MD) is the highest peak demand of all the daily peak demands over a month. Battery energy storage system (BESS) can be used and controlled appropriately to reduce the monthly maximum demands such that customers can reduce their maximum demand charges and power utility companies can defer their network reinforcement and augmentation of generating capacity. It is important to ensure that any technique employed for BESS to reduce MD is consistent and reliable in reducing every single daily peak demand to deliver the financial benefits. An online deep learning regression controller is a new control technique introduced in this paper that is different from conventional methods. This technique operates using its regression ability to generate the best possible threshold in real-time such that BESS supplies real power to the load when the load demand exceeds the threshold. This regression ability is absent in all the published BESS controllers. Its regression ability is trained with historical data and actively updated with its new data by using its online learning feature to capture any long-term changes in the demand due to the human behavioural or climate changes. The performance of the controller is evaluated through simulation and real-world experiments. The simulation studies show that the controller achieves improved performance when benchmarked against other deep learning models and existing controllers with an average maximum demand reduction factor of 75.08%. The proposed controller achieves about 10.89% higher maximum demand reduction than the next closest contender. Furthermore, the experimental results validate its effectiveness, showing an average maximum demand reduction factor of 73.49%. Overall, the proposed controller effectively reduces maximum demand in both simulation and real-world scenarios, achieving an improvement of 175.42% compared to a battery system with basic scheduling. © 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.en-USBattery energy storage system (BESS)Deep learningMaximum demand reductionOnline learningBehavioral researchData miningDigital storageE-learningElectric batteriesEnergy storageHuman computer interactionLearning systemsRegression analysisBattery energy storage systemBattery energy storage systemsDemand reductionMaximum demandPeak demandPerformanceReal- timeControllerstext::journal::journal article