Multi-objective optimization is essential in complex systems, where multiple conflicting objectives must be balanced simultaneously. Traditional multi-objective evolutionary algorithms (MOEAs) are effective in approximating the Pareto front but typically rely on iterative population-based optimization, which may limit their applicability to online decision-making. Reinforcement learning (RL) offers an alternative, while adaptive trade-off selection during policy execution remains an important issue in multi-objective decision-making. This study proposes a multi-objective deep RL framework (MODRL) to address these issues. A scalarization-based reward design enables policy learning under varying trade-offs, while augmenting the state with weight vectors allows the policy to internalize preference information and generalize across different preference settings. In addition, a Q-value-guided dynamic weight adaptation mechanism is employed during inference, enabling adaptive adjustment of trade-offs according to system conditions. Experimental results on the ZDT and WFG benchmark suites show that MODRL achieves competitive optimization performance compared with representative MOEAs, while enabling efficient online policy execution after offline training. Furthermore, the effectiveness of the proposed framework is validated in the Benchmark Simulation Model No. 1 (BSM1) wastewater treatment process, demonstrating its potential for practical multi-objective optimization and decision-making. Note to Practitioners—This paper presents a reinforcement learning framework for multi-objective operational decision-making, demonstrated on a wastewater treatment process where conflicting objectives such as operating cost and effluent quality must be balanced. The framework learns decision policies directly from interaction with a process simulator, which is useful when accurate mechanistic models are difficult to obtain. During online operation, the learned policy can adapt objective trade-offs according to current system conditions. In practice, engineers need a reliable simulator or digital twin for offline training, as well as clearly defined performance indicators and operational constraints. The method is flexible and can incorporate different deep reinforcement learning algorithms according to practical requirements.