Deep Learning‐Based Intelligent Resource Management for Renewable Energy Engineering Projects

Abstract

This study integrates deep reinforcement learning (DRL) with lean management for renewable energy project scheduling. The problem is formulated as a constrained Markov decision process with weather‑dependent dynamics and a composite reward such as schedule, cost, quality, and constraints. A four‑layer architecture (convolutional, recurrent, and attention layers) and a masked proximal policy optimization actor–critic enhanced with prioritized experience replay and meta‐learning learn adaptive policies; four lean principles map directly onto the DRL loop. A three‐stage feasibility filter—hard action masking, soft resource rescaling, and penalty fallback—ensures that every executed decision satisfies project‐network, resource, weather‐window, grid, and environmental constraints. Validation on 186 Northwest China projects shows 33.0% shorter duration, 29.2% lower cost, and 87.2% resource utilization while preserving quality. Compared with eight state‐of‐the‐art deep‐learning and classical baselines, the proposed approach delivers an average improvement of approximately 2.9 percentage points in schedule efficiency over the strongest deep‑learning baselines, and reduces mean absolute error by approximately 41%–50% on large‑scale projects (>200 activities). The proposed method retains ∼80% performance under 30% noise and ≥97.8% on CPU‑only hardware, supporting practical adoption by small‐to‐medium enterprises and on‐site project offices.