Cash collection for 1,630 terminals

Task

Approach

1. 01Forecast every terminal

   One N-BEATS model per terminal, 1,630 in all: 14 days in, 15 days out, trained jointly on the terminal’s own deposits and on the average of its five nearest neighbors. The median error on the hold-out weeks was about 9 thousand rubles a day.

2. 02Turn forecasts into deadlines

   Each morning every terminal gets its days left: zero if it is already over the cap or at the 14-day limit, otherwise the first day on which forecast deposits would push it over the cap.

   $${\delta }_t=\min (14-(d-L_t),\min \{k:B_t+\sum _{j=1}^k{\hat{f}}_{t,d+j}\ge C\left\}\right)$$

3. 03A pool of routes

   Routes are nearest-neighbor chains that fill a 12-hour shift: one seeded at every terminal and computed once, plus fresh chains through today’s must-visit terminals, so every day has a feasible cover.

4. 04Choose today’s routes

   A set-covering MIP picks at most K routes. Must-visit terminals are covered at least once, the rest at most once, and optional visits are weighted by how crowded their deadline group is. That is a cheap look-ahead that spreads work across days.

   $$\max \sum _{t:{\delta }_t>0}\frac{G_{{\delta }_t}}{{\delta }_t}\sum _r{a}_{tr}{x}_r\text{s.t.}\sum _r{a}_{tr}{x}_r\ge 1({\delta }_t=0),\sum _r{a}_{tr}{x}_r\le 1({\delta }_t>0),\sum _r{x}_r\le K$$

5. 05Three months in a loop, then the smallest fleet

   Each day’s plan is applied to actual deposits, deadlines are recomputed from fresh forecasts, and the loop runs through all 91 days. The fleet is reduced until some day becomes infeasible. A separate covering model with perfect information estimates how far it could go.

   $$\min K\text{s.t.}\sum _{d\in I}\sum _r{a}_{tr}{x}_{rd}\ge 1\forall (t,I)\in \mathcal{I},\sum _r{x}_{rd}\le K\forall d$$

Result

Stack