Photovoltaic capacity planning has shifted from single-asset nameplate selection to integrated energy system (IES) sizing, where reinforcement learning (RL) with a multistage decision strategy beat mixed-integer linear programming (MILP) and multi-objective optimization (MOO) on an industrial-park case in a July 2023 Springer Energy Systems study [S1].
The macro pipeline is now measured in terawatts, not gigawatts: the IEA projects global module production capacity will exceed 1.5 TW by 2035, while China had already run 7 consecutive years of leading new installed PV capacity and 13 consecutive years leading module production as of end-2019 [S2][S8].
Decision Criteria: Economy, Environment, Efficiency, Safety
The RL framework in the Springer study ranks candidate capacity configurations against four explicit criteria — economy, environment, efficiency, and safety — and introduces new indicators including a capacity factor criterion, an installed capacity ratio of renewable devices, a generation ratio, and an adjust rate of energy generators [S1]. For an industrial-park test case, the RL multistage policy reduced action dimensionality and outperformed MILP/MOO on the optimal capacity configuration, with sensitivity analysis confirming robustness of the recommended mix [S1].
In hybrid microgrids, a 2023 ScienceDirect study used a Salp Swarm Algorithm (SSA) metaheuristic to co-optimize capacity planning and energy management for a standalone PV–wind–biomass–hydrogen–battery system, evaluating trade-offs against multiple objective functions to find the optimum capacity split [S3]. The result is a planning toolset where SSA, RL, MILP, and MOO all converge on the same insight: PV sizing is no longer decided in MW alone but as a ratio against storage, load, and demand-side flexibility. For process plants weighing capex against kWh landed cost, this is the metric shift that matters — see the broader selection logic on industrial valves for analogous trade-off mapping in adjacent process equipment.
Production-Capacity Backdrop: Polysilicon, Modules, Wafer Cost
China's polysilicon output reached roughly 290,000 tonnes in the first three quarters of 2020, up 18.9% year-on-year, with module production capacity clearing 80 GW (+6.7% YoY) in the same window [S2]. LONGi's Li Zhenguo reported the silicon-wafer price had fallen to 3 CNY from ~100 CNY a decade earlier, and module price to 1.7 CNY/W from 30 CNY/W over the same period, with PV generation cost below 0.1 CNY/kWh in high-irradiance sites [S2].
For capex planners, the cost curve dictates that PV now competes with wind on LCOE in many jurisdictions, and capacity-planning models must reflect that PV modules are no longer the dominant cost — balance-of-system (BOS), inverters, flow meters for thermal integration, and storage round out the bill of materials. Module manufacturing equipment selection itself is now a spec-driven activity tied to throughput per gigawatt; the PV manufacturing equipment selection map walks cell, module, and test-line criteria against the same terawatt-scale backdrop [S4][S8].
Who This Planning Approach Is For — and Where It Breaks
RL-based and SSA-based capacity planning is built for industrial-park IES operators, microgrid developers, and hybrid PV–wind–biomass–hydrogen–battery project owners who need to size multiple devices simultaneously against economic, environmental, efficiency, and safety criteria [S1][S3]. A separate State Grid Zaozhuang / Shandong Guorui study, published 2018, used time-series production simulation for distribution-network PV installed-capacity planning under State Grid Science & Technology project 2017A82, indicating the methodology is mature enough to have entered utility planning practice in China for several years [S9].
Where it does not fit: a single rooftop PV retrofit on a commercial building, where the optimum is a 1:1 replacement and the cost of running an RL or SSA optimization exceeds the savings; or a greenfield utility-scale PV farm in Spain that simply needs a nameplate matching a PPA, where Spain's installed-PV dataset of 2.8 GW (2021) and ~4 GW (2019) shows the market is small enough that simple load-match heuristics still dominate [S6]. The Chinese policy framing published in People's Daily Online in May 2024 rejects the "overcapacity" narrative and treats supply-demand equilibrium as the governing signal, which is the macro posture planners should price into long-horizon capacity models rather than tariff-curve fads [S7].
Standards, Verification, and Plant-Side Integration
Module and cell quality gates are governed by IEC and UL standards; the Solar Panel Manufacturing Quality Standards map consolidates the IEC/UL/ISO gating stack that any capacity-planning model should treat as a hard constraint on equipment eligibility [S4][S8]. For the smart-factory side, Industry 4.0 metamodels and IIoT stacks feed back into the capacity-planning loop with real OEE, which is where planning stops being an annual exercise and becomes a continuous control problem, as detailed in the Solar Panel Industry 4.0 reference [S4].
For hazardous-area PV-plus-BESS sites (hydrogen storage adjacent to PV inverters, for example), zone mapping and Ex d/Ex e/Ex i selection feed into the same capacity model, and the explosion-proof installation reference is the practical check before the design freeze. Plant-side, the same pressure transmitter and PLC selection logic that governs a process skid applies to PV-hybrid skids where hydrogen compression, battery thermal management, and grid-tie inverters all need closed-loop control.
Comparison: Planning Methods Against Four Decision Criteria
For a project team picking a capacity-planning method, four criteria separate the candidates: solution quality on multi-criteria problems, computational cost / dimensionality handling, ability to co-optimize with energy management, and track record on industrial-park validation. RL with multistage decision policy and multi-criteria evaluation handles all four: it was demonstrated to outperform MILP and MOO on the Springer industrial-park case across economy / environment / efficiency / safety indicators, with reduced action dimensionality improving computational efficiency [S1]. SSA metaheuristic (ScienceDirect, 2023) handles multi-criteria + co-optimization with energy management for PV–wind–biomass–hydrogen–battery, with strong performance reported for the standalone hybrid case [S3]. Time-series production simulation (State Grid, 2018) is the lowest-computational-cost option and is suited to distribution-network PV integration where the planning horizon is the binding constraint [S9]. MILP remains the incumbent baseline for tractable, convex cases but loses to RL on the Springer benchmark when criteria count and action dimensionality grow [S1]. For most 2026 projects, the practical answer is RL or SSA on the optimization side with time-series simulation as a sanity check.
Limits, Failure Modes, and Data Hygiene
RL-based planning assumes the action space and reward function faithfully represent the project; if the installed-capacity ratio and adjust-rate indicators are mis-weighted, the optimizer will overbuild flexibility assets (battery or hydrogen) at the expense of PV [S1]. SSA convergence on hybrid systems depends on the penalty factors in the objective function — without them, the algorithm over-allocates to whichever device has the cheapest per-kW nameplate, which today is still PV [S3]. Time-series production simulation fails outside its calibrated load and irradiance profile; transplanting the State Grid Zaozhuang methodology to a tropical island microgrid without re-fitting is a known footgun [S9]. And macro forecasts that pin a 1.5 TW global module nameplate by 2035 are scenario-dependent; the IEA report itself frames it as a projection, not a commitment, so any capacity plan that uses it as a hard input should carry a sensitivity band [S8].
Trackable signals for the next planning cycle: the 2025 monocrystalline-silicon production-capacity dataset published by CBCIE, which sets the input-cost baseline for any 2026 PV capex model [S4]; the EGING PV (SH600537) capacity disclosures as a listed-A-share module producer with 2,129.46 million CNY registered capital, which act as a real-economy capacity marker [S5]; and the People's Daily Online editorial line that continued into 2024 rejecting the "overcapacity" framing, which signals that domestic Chinese policy will not impose nameplate cuts on PV manufacturing in the near term [S7].