Warehouse Storage Location Recommendation: Architecture, Recall, and Ranking Strategies

This article describes a recommendation solution for the DeWu Warehouse Management System (WMS) that helps decide where to place items on shelves. Optimizing the shelving process is crucial because the initial placement influences all downstream warehouse operations such as picking and packing.

1. System Overview

The overall workflow of the WMS is illustrated in the diagram below.

1.1 Industrial Recommendation Architecture

The system is divided into three parts: online, near‑line, and offline. The online part performs recall, optional coarse ranking, and fine ranking before applying business policies. The near‑line part collects real‑time user feedback and updates the model. The offline part stores click logs, prepares training data, and periodically retrains the model using a distributed machine‑learning platform.

1.2 Challenges for Warehouse Slot Recommendation

Massive SKU variety and quantity make manual placement infeasible.

Balancing space utilization versus retrieval efficiency.

Business constraints such as category‑specific zones and slot capacity limits.

Operational issues like uneven inventory distribution and random shelving by staff.

Two recommendation schemes are proposed.

2. Slot Recommendation Schemes

2.1 Recall Engine

The recall stage quickly filters a large pool of slots to a manageable candidate set. It consists of four modules: label index construction, recall, weighting‑balanced ranking, and the recommendation model.

{
  "仓库_库区": ["完全限制", "sku限制", "货主限制", "质量等级+业务类型限制", "无限制"]
}

{
  "仓库_库区_完全限制": ["A_B_6_100", "B_B_6_100", "*"],
  "仓库_库区_sku限制": ["SKU1", "SKU2", "SKU3", "*"],
  "仓库_库区_货主限制": ["小A", "小B", "小C", "*"],
  "仓库_库区_质量等级_业务类型": ["100_6", "200_7", "100_1", "*_3", "100_*", "*"],
  "仓库_库区_无限制": ["库位A", "库位B", "库位C"]
}

{
  "仓库_库区_部分_skuidA_货主A_bizTypeA": ["库位A", "库位B", "库位C"]
}

{
  "仓库_库区_AB010101": {
    "locationCode": "AB010101",
    "limitCode": 8,
    "currentQty": 20,
    "bizType": [200,100,900],
    "skuList": ["123","12355"],
    "maxCommodityQty": 8,
    "warehouseCode": "仓库",
    "aeraCode": "AB",
    "upperSeq": "1"
  }
}

2.2 Ranking Engine

The ranking stage assigns scores to the recalled slots from multiple perspectives (space usage, time cost, sales potential) and selects the highest‑scoring slot.

Weight‑balancing example:

Score = 50*(space/100) + 30*(time/100) + 0*(sales/100)

Maximum space utilization formula:

Utilization = (currentCapacity + incomingQty) / maxCapacity

Optimal path cost:

S = aP + bQ + c(max(R,2)-2)
T = (aP+bQ)/vx + c(max(R,2)-2)/vy

Sales score is derived from recent inbound/outbound volumes and projected demand.

3. Summary

The proposed recall‑and‑ranking framework helps DeWu’s warehouse managers place items efficiently, reduce picking time, and lower inventory costs. For rapidly changing data, Redis is recommended for the label index; for more stable, large‑scale scenarios, Elasticsearch may be used.