Design and Evaluation of an AI-Driven Digital Twin Platform for Battery Energy Storage System Assembly

Sammanfattning

Background: Battery Energy Storage Systems (BESS) are important for renewable energy integration, grid flexibility, and reliable energy management. A BESS installation is a complex assembly of interconnected components such as battery modules, switches, inverters, cables, sensors, and control devices. Missing, misplaced, or incorrectly connected components can create safety risks, reduce reliability, and increase troubleshooting effort. Current validation workflows often rely on static manuals, drawings, and manual inspection, which may not fully reflect the real-time site conditions. This creates a gap between expected configurations and actual system states, making fault identification difficult and error-prone. Objectives: This thesis develops and evaluates an AI-driven Digital Twin platform for validation and troubleshooting support in BESS assemblies. The platform operates through a three-stage pipeline: (1) YOLOv8-based Computer Vision (CV) detects visible site components, (2) observed data is compared with the expected system configuration through Static-Dynamic Digital Twin Graph (DTG) comparison, and (3) a deterministic rule engine analyzes discrepancies and generates actionable corrective instructions. The output is delivered through an Augmented Reality (AR) interface, enabling intuitive, context-aware guidance for technicians. Methods: The system was developed using Design Science Research. Evaluation was conducted through controlled technical experiments using real dataset images, generated Dynamic DTGs, Static-Dynamic graph comparison, deterministic rule-engine outputs, RAG retrieval records, and selected live local LLM checks. The evaluation included Stage-I detection tests, Stage-II graph comparison tests, Stage-III reasoning tests, fourteen controlled fault-injection scenarios, an end-to-end image-to-instruction combination test, failure checks, runtime measurements, and a guarded live LLM evidence suite using llama3:latest. Results: The final YOLOv8 model reached a Mean Average Precision at an IoU threshold of 0.50 (mAP@50) of 0.888 and a stricter averaged Mean Average Precision across IoU thresholds from 0.50 to 0.95 (mAP@50-95) of 0.776 on the validation data. The expanded Stage-I image evidence used nine representative images and produced 53 detections across 14 detected classes. The controlled scenario matrix included fourteen injected discrepancy scenarios across eight fault clusters; within this controlled test set, the predefined scenario checks matched the expected outputs. The end-to-end combination test processed ten images, produced 56 YOLO detections, generated 23 Stage-II comparison records, and produced 37 deterministic troubleshooting instructions. Stage-II graph comparison and deterministic Stage-III reasoning completed within milliseconds, while live LLM enhancement was slower and therefore better suited for asynchronous wording refinement. Conclusions: The results support the prototype-level feasibility of an AI-driven Digital Twin framework that combines CV, graph-based comparison, deterministic reasoning, RAG-supported instruction wording, and AR-assisted interaction for BESS validation and troubleshooting. The prototype detected the controlled configuration differences used in the evaluation, generated traceable troubleshooting instructions, and avoided unnecessary repair actions in nominal controlled cases. The evaluation should be interpreted as technical system validation under controlled conditions, not as proof of full field robustness or improved technician performance. Large-scale field testing and human-centered usability studies remain future work.