交通运输工程与信息学报

【背景】链路观测有限使OD估计具有高度欠定性，日间波动、异常观测及先验偏差易放大噪声并诱发OD结构漂移，降低估计结果的稳定性与结构可解释性。【目标】面向多日观测条件，实现平均—日度OD协同估计，并稳健识别关键OD结构。【方法】构建准稀疏两阶段随机规划OD估计模型(SP-QSOD)，以跨日共享的平均OD为第一阶段结构锚点，以日度OD为第二阶段追索变量，联合利用K日观测；通过对先验修正、日间偏移和观测残差施加三类L₁约束，并采用样本平均近似将模型转化为确定性线性规划求解。【数据】在SiouxFalls路网构建基准、重尾异常、映射误差、先验偏移和小样本五类场景，并与普通最小二乘(OLS)、广义最小二乘(GLS)及确定性准稀疏OD估计(D-QSOD)进行对比。【结果】相对先验OD，SP-QSOD在映射误差(S2)、先验偏移(S3)与小样本(S4)场景下的平均OD RMSE改善率分别为21.5%、22.5%和16.8%；以非显著OD为正类时，对应的F1_non-sig分别为0.80、0.82和0.84，五类场景(S0–S4)下Accuracy稳定在0.71–0.75。扩展实验表明，该方法在更一般映射设定下仍保持较稳定的结构识别表现，并可迁移至更大规模路网。【结论】两阶段均值锚定—日度追索框架与准稀疏正则的耦合，可在映射误设、先验偏移与小样本等条件下抑制噪声放大，在估计精度与结构可解释性之间取得稳健平衡。【应用】适用于检测器布设稀疏且先验OD或OD–链路映射可能存在偏差条件下的OD推断与交通规划评估。

[Background] Limited link observations make origin-destination (OD) estimation highly underdetermined. Day-to-day fluctuations， anomalous measurements， and prior bias may amplify noise and induce structural drift in OD patterns， thereby weakening estimation stability and structural interpretability. [Objective] This study aims to achieve coordinated estimation of mean and daily OD demands under multi-day observations while robustly identifying key OD structures. [Method] A quasi-sparse two-stage stochastic programming model for OD estimation (SP-QSOD) is proposed， where the cross-day shared mean OD serves as the first-stage structural anchor and the daily OD vectors are treated as second-stage recourse variables. Multi-day observations are jointly utilized， and three L₁-norm penalties are imposed on prior correction， inter-day deviation， and observation residuals. The model is solved as a deterministic linear program via sample average approximation. [Data] Five scenarios， including baseline， heavy-tailed anomaly， mapping error， prior shift， and small sample， are constructed on the SiouxFalls network. The proposed method is compared with ordinary least squares (OLS)， generalized least squares (GLS)， and deterministic quasi-sparse OD estimation (D-QSOD). [Results] Relative to the prior OD， SP-QSOD improves the mean-OD RMSE by 21.5%， 22.5%， and 16.8% under the mapping-error (S2)， prior-shift (S3)， and small-sample (S4) scenarios， respectively. When non-significant OD pairs are treated as the positive class， the corresponding F1_non-sig scores are 0.80， 0.82， and 0.84， while the classification accuracy remains within 0.71–0.75 across all five scenarios (S0–S4). Additional experiments indicate that the proposed method maintains relatively stable structural identification performance under a more general mapping setting and can be extended to a larger-scale network. [Conclusion] By coupling a two-stage mean-anchoring and daily-recourse framework with quasi-sparse regularization， the proposed method suppresses noise amplification under mapping misspecification， prior shift， and limited samples， and achieves a robust balance between estimation accuracy and structural interpretability. [Application] The proposed method is applicable to OD inference and transport planning evaluation in settings with sparse detectors and potential bias in prior OD information or OD-link mappings.