ATM Cash Cassette Replenishment Using Ensemble Demand Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing manual forecasting methods for ATM cash demand are inaccurate and inefficient, leading to ATMs running out of cash or holding excess cash, resulting in increased operating costs and unnecessary refill trips.
Innovation Solution
A machine learning-based system using an ensemble model approach with a fully connected neural network and gradient boosting framework processes real-time ATM sensor data to predict cash demand, optimizing cash distribution and reducing the number of replenishment trips.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual forecasting methods are used to predict cash demand, then the process is simple to operate, but the prediction accuracy is low leading to ATMs running out of cash or holding excess cash
Solution Approach 1:
The patent replaces manual forecasting methods with an automated machine learning system that processes sensor data from ATMs. The system uses ensemble models (neural networks and gradient boosting) to automatically predict cash demand, substituting human judgment with computational algorithms that can analyze multiple factors simultaneously including seasonality, holidays, public events, and recent withdrawal trends.
Solution Approach 2:
The system enables self-service by allowing the forecasting system to automatically update itself through continuous learning from new data. The machine learning models are retrained periodically with fresh sensor data, enabling the system to adapt and improve its predictions autonomously without requiring manual intervention or updates to the forecasting logic.
2Reliability
If ATMs hold excess cash to prevent running out, then cash availability is improved, but operating costs increase due to unnecessary refill trips
Solution Approach 1:
The system performs preliminary action by predicting future cash demand before the ATM actually runs out of cash. The forecasting model generates predictions that allow the replenishment system to schedule deliveries in advance, optimizing the timing and quantity of cash deliveries to match actual demand patterns rather than reacting to emergency situations.
Solution Approach 2:
The system implements feedback by continuously monitoring actual cash transactions and sensor data from ATMs, comparing predicted versus actual cash demand, and using this information to retrain and improve the forecasting models. This closed-loop feedback mechanism ensures the system learns from past performance and continuously optimizes cash availability while minimizing unnecessary trips.
3Reliability
If emergency refills are performed when ATMs run out of cash, then cash availability is restored, but lead times extend up to 36 hours impacting service
Solution Approach 1:
The system performs preliminary action by predicting cash depletion events before they occur and scheduling replenishment deliveries in advance. The forecasting model identifies ATMs that will run out of cash within a certain time horizon, allowing the logistics system to arrange deliveries proactively rather than reactively, thereby eliminating emergency refill situations and reducing lead times.
4Reliability
If more frequent replenishment trips are made, then cash availability is improved, but the number of trips and associated costs increase
Solution Approach 1:
The system applies parameter changes by dynamically adjusting replenishment quantities and frequencies based on predicted demand parameters. Instead of using fixed replenishment schedules or quantities, the system modifies these parameters according to forecasted cash withdrawal patterns, seasonal variations, and local factors, optimizing the balance between cash availability and trip efficiency.
Data Source
AI summary
A specific architecture is proposed that utilizes two models being operated in parallel as an ensemble model approach based on Applicant's testing with physical machines. The ensemble model approach is provided as a physical system that operates two models simultaneously, both models being trained as candidate models. Both models are utilized during inference time separately to optimize a loss function (e.g., MAE performance), and during inference, the model with a superior MAE performance is used to control ATM replenishment control signal generation. The two models being used together include a first model, a fully connected neural network data architecture, and a second model, a tree-based learning algorithm provided as a gradient boosting framework (e.g., the Light Gradient-Boosting Machine, also known as the LightGBM). From a practical perspective, the ensemble models can be operated with a prediction buffer configured to allow for specific parameter tuning.