Optimal Real-Time Control of Combined Sewer Networks

Authors: M. Papageorgiou, R. Mayr Year: 1985 Tags: combined-sewer-networks, real-time-optimal-control, discrete-maximum-principle, stormwater-management, urban-drainage, repetitive-optimization

TL;DR

Formulates a discrete-time optimal control problem for combined sewer networks that minimizes flooding and overflow by coordinating detention reservoirs, solved via the Davidon-Fletcher-Powell (DFP) gradient algorithm applied to the discrete maximum principle. Tests sensitivity to weighting factors, sample time, prediction accuracy, and optimization horizon on 11 synthetic inflow events for a single three-reservoir example network, showing consistent superiority over a hand-crafted heuristic.

First pass — the five C's

Category. Research prototype — applied optimal control / water resources control.

Context. Urban stormwater / sewer control subfield; builds directly on Papageorgiou (1983) ASCE J. Env. Eng. (multilayer control structure and full algorithm derivation) and Papageorgiou (1984) IFAC World Congress (multilayer sewer network control). Simplification validity checked against Wanka and Koniger (1984) HVM hydrograph-volume method. Rainfall prediction context from Schilling and Okroy (1981).

Correctness. Three load-bearing assumptions: (1) a simplified conservation-equation model (storage, delay, diversion elements) adequately represents the true hydraulic behavior for control purposes — partially verified via HVM comparison; (2) DFP finds solutions close enough to the global optimum despite nonconvexity introduced by the diversion element (eq. 5–6) — not proven, acknowledged as unguaranteed; (3) 11 hypothetical events sufficiently span real operational diversity — unverified.

Contributions. - Five-term weighted performance criterion enabling explicit priority ordering of flood avoidance, overflow minimization, treatment-plant utilization, storage distribution, and flow smoothness. - Quantified sensitivity of flooding and overflow to sample time T (T=10 min → 10% overflow increase; T=20 min → substantial deterioration vs. T=5 min baseline). - Demonstration, via HVM cross-validation, that simple storage-element and time-delay approximations are adequate for control-layer optimization. - Repetitive optimization scheme with an explicit inflow prolongation rule enabling real-time application under uncertain predictions.

Clarity. Concise for a conference paper but explicitly defers full results to a future paper, leaving key ablations (e.g., per-event breakdown, full HVM comparison tables) unreported.

Second pass — content

Main thrust: A DFP-based optimal control strategy operating on a simplified sewer network model outperforms a sophisticated hand-crafted heuristic by 42% on average (flooding + overflow) across 11 synthetic events, runs in ~60 CPU-seconds per horizon on a DEC-2060, and degrades by only 22% under 15-minute accurate predictions (kp=3) versus perfect foresight.

Supporting evidence: - Heuristic vs. optimal: average 42% increase in combined flooding and overflow across 11 events; heuristic never outperformed optimal on any event. - Inaccurate predictions (kp=3, i.e., accurate for next 15 min): 22% average increase in flooding + overflow relative to perfect-prediction optimal. - No predictions (kp=0) vs. kp=3: only an additional 8% average increase in flooding + overflow, indicating method is viable without a rainfall-runoff model. - Optimization horizon K<5 intervals (< 25 min) yields "extremely bad" results; adequate horizon is 15–35 intervals (75–175 min) depending on current inflow. - Computation: ~60 CPU-seconds per run (K=60, T=5 min) on DEC-2060; no convergence failures reported.

Figures & tables: Fig. 4 shows outflow trajectories for T=5 vs. T=20 min (axes labeled, single deterministic scenario, no uncertainty bands). Figs. 5–7 illustrate a dual-peak event: inflow hydrographs, storage element volumes, and control trajectories (axes labeled with time in hours and flow in m³/s; no error bars). Fig. 8 illustrates the repetitive optimization inflow-updating mechanism (real vs. predicted inflows at successive re-optimization steps, clearly annotated). No tables are present; no statistical significance or confidence intervals are reported anywhere.

Follow-up references: - Papageorgiou (1983), ASCE J. Env. Eng. — full derivation of the multilayer structure and optimization algorithm; necessary to understand the theoretical foundation. - Papageorgiou & Mayr (1986, submitted) — algorithm comparison motivating DFP selection; needed to evaluate solver choice. - Wanka & Koniger (1984), HVM method — the hydraulic simulator used to validate model simplifications; relevant for assessing fidelity claims. - Schilling & Okroy (1981) — rainfall prediction basis for kp parameterization; relevant for real-world deployment.

Third pass — critique

Implicit assumptions: - A single three-reservoir synthetic network is taken as representative of general combined sewer networks; topology, scale, and delay structure are not varied. - The 11 synthetic inflow events are assumed to span operationally relevant scenarios; no statistical characterization of the event space is provided. - The multilayer structure's direct-control (feedback) layer is assumed capable of tracking optimal trajectories closely; this layer is entirely out of scope and never evaluated. - Re-optimization period k (how often the K-step problem is re-solved) is not defined explicitly; thus the claimed 60-second feasibility cannot be verified against actual timing constraints. - DFP convergence to a near-global optimum is assumed adequate in practice; no analysis of how far local solutions deviate from global ones.

Missing context or citations: - No engagement with model-predictive control (MPC) literature from the control systems community, which addresses the same receding-horizon structure more formally. - No comparison with simpler threshold-based or rule-based strategies actually used in practice at the time; the only baseline is an ad hoc heuristic designed by the same authors. - No citation of alternative hydraulic modeling approaches for the optimization layer beyond the authors' own prior work.

Possible experimental / analytical issues: - The heuristic strategy was derived by the authors themselves after inspecting optimal results, creating clear risk of anchoring bias; an independently designed or practitioner-supplied heuristic would be a stronger baseline. - The 42% improvement figure is an average with no variance, range, or per-event data reported, making it impossible to assess robustness or identify failure modes. - Global optimality is unguaranteed due to the nonlinear diversion element (eqs. 5–6); no sensitivity to initialization or frequency of suboptimal convergence is reported. - The inflow prolongation rule beyond kp is acknowledged as "rather arbitrary," yet no sensitivity analysis of alternative prolongation rules is performed. - All inflow events are hypothetical; no real rainfall or runoff measurements are used, so generalization to actual catchment behavior is undemonstrated. - HVM cross-validation results are summarized qualitatively ("reasonable," "sufficient") with no quantitative difference metrics reported.

Ideas for future work: - Apply the method to a real sewer network using measured storm events to test whether the 42% improvement over heuristics and the 22%/8% prediction-error penalties hold outside synthetic conditions. - Systematically vary the inflow prolongation rule (e.g., climatological persistence, nowcast extrapolation) and measure sensitivity of flooding and overflow to rule choice. - Investigate frequency and magnitude of local-vs-global optimum discrepancies by multistart DFP runs, and assess whether a convex relaxation of the diversion element is tractable. - Evaluate the full multilayer system (optimization + adaptation + direct-control layers together) to determine how tracking error in the lower layer degrades performance guarantees established here.