Implementation and validation of a deep learning-based predictive dispatch algorithm for off-grid PV/diesel/battery hybrid power systems

This study presents the development and lab-scale implementation of an innovative deep learning-based predictive dispatch algorithm for off-grid PV/diesel/battery hybrid power systems aimed at improving energy reliability in rural communities. The algorithm uses long short-term memory (LSTM) models to forecast solar irradiance and electrical load and applies this information for real-time energy management. Unlike conventional Cycle Charging (CC) and Load Following (LF) strategies, the predictive algorithm dynamically determines when to start or stop the diesel generator, when to charge or discharge the battery, and how to balance PV output to minimize losses and maximize renewable utilization, while also optimizing the battery’s state of charge to prevent overcharging or deep discharging. By forecasting both solar availability and load demand, the algorithm can proactively schedule diesel operation during periods of low irradiance, allocate PV energy efficiently for immediate use or storage, and reduce unnecessary generator cycling, thereby extending equipment lifetime and improving overall system efficiency.

The algorithm was integrated into an Intel Mini-PC controller and tested with a prototype off-grid system comprising a 1.44 kWp PV array using Sharp ND-Q245F7 modules, a 5 kWh Dyness DL5.0C lithium battery, and a 5 kW Deye SUN‑5K‑SG05LP1 hybrid inverter. Device communication and real-time control were implemented via Modbus and MQTT, with inputs from weather sensors and digital power meters. Figure 1 shows the system integration and control architecture, illustrating how the AI core exchanges data with inverters, sensors, and diesel control loops to manage dispatch decisions in the off-grid setup. Figure 1 shows the system integration and control architecture, illustrating how the AI core exchanges data with inverters, sensors, and diesel control loops within the off-grid setup. Laboratory-scale tests showed that the predictive strategy reduced diesel runtime by up to 45%, cut PV clipping losses by 40%, and achieved a performance ratio of 75%, significantly higher than CC (56%) and LF (64%). The control process also improved battery health by avoiding overcharging and deep discharging, while cutting diesel fuel consumption and associated CO₂ emissions.

From Table 1, it was found that the predictive dispatch algorithm reduced diesel fuel use by up to 45% and clipping loss by around 40% compared with Cycle Charging and Load Following, achieving a performance ratio close to 75%. Together, these results confirm that the predictive control algorithm can be implemented and operated effectively in off-grid hybrid setups and demonstrate its strong potential for wider application in electrifying remote national parks and isolated communities. This research was financially supported by the National Research Council of Thailand (NRCT).