Business and operations
Machine learning for operations
Operations teams use machine learning to forecast demand for staff and fleet capacity, estimate delivery and service times, and predict which equipment is closest to failure. A planning or optimisation step then turns those estimates into schedules and maintenance lists.
How machine learning is used in operations
Forecasting demand for capacity
Models forecast orders, calls or trips per hour and location so that staff, vehicles and stock can be planned in advance. Probabilistic forecasts give a range, which lets planners choose how much buffer to hold. Seasonal statistical models and recurrent or tree-based models trained across many related series are used.
Real exampleDeepAR (Amazon)DeepAR trains an autoregressive recurrent network across many related time series to produce probabilistic forecasts, with retail demand forecasting as the motivating case.
Estimating travel and service times
Arrival-time models predict how long a trip, job or service visit will take from the route, time of day and recent traffic. Schedulers then use those estimates instead of fixed averages. Gradient-boosted trees are common, and graph neural networks are used for road networks.
Real exampleGoogle Maps ETA prediction with graph neural networksThe paper presents a graph neural network estimator for estimated time of arrival that is deployed in production at Google Maps.
Predicting equipment failure
Predictive maintenance models use sensor readings such as temperature, pressure and vibration to estimate remaining useful life or the chance of failure soon. Maintenance teams then service the machines most at risk first. Regression on rolling sensor features and sequence models are typical.
Real exampleNASA Prognostics Center of Excellence data repositoryNASA's repository publishes run-to-failure prognostic data sets, including turbofan engine degradation simulated with C-MAPSS under different operating conditions and fault modes.
Controlling energy-intensive systems
Models trained on historical sensor data predict how a plant or building will respond to setting changes, and a control step picks settings that use less energy while staying within limits. Neural networks are common for these multi-sensor problems.
Real exampleDeepMind data centre coolingDeepMind trained ensembles of deep neural networks on historical data-centre sensor data to predict future power usage effectiveness and recommend cooling actions.
Can you learn machine learning for operations?
Yes. Operations analysts and engineers already understand the processes and constraints, and the modelling builds on regression and forecasting they may have used in spreadsheets. Learn Python and pandas, then statistics, then regression, time series and anomaly detection, and finally connect predictions to an optimisation step.
- Python and pandas for operational dataLoad event logs, sensor exports and order histories, resample them to hours or days and build lagged features.Learn it in Data Tools
- Statistics for variabilityDistributions, quantiles and estimation explain why planning to the average forecast causes shortages half the time.Learn it in Statistics
- Regression, time series and anomaly detectionForecast demand, estimate durations and flag unusual sensor readings, always compared with the rule the team uses today.Learn it in Machine Learning
- Optimisation basicsLearn linear programming and simple scheduling with a solver such as OR-Tools, so a forecast can become a rota or route plan.
Moving into machine learning for operations
If you already work in the field
Operations managers and industrial engineers already know the constraints, costs and data sources that decide whether a model is useful. Add Python, forecasting and honest validation on later periods. A realistic first move is to forecast one queue, line or depot's demand and compare the plan it produces with the current one.
If you come from data science or ML
If you come from ML, learn how the operation actually runs: shift patterns, service levels, lead times and what a planner can change on the day. Shadow a planner or maintenance team, because the decision a model feeds sets the right metric.
Roles that do this workOperations research analystSupply chain data scientistDemand plannerReliability engineerIndustrial engineer
Tools you will use
| Tool | What you use it for |
|---|---|
| scikit-learn | Fit demand, duration and failure-risk models and compare them with simple rules. |
| StatsForecast | Forecast hourly or daily demand across many sites with fast statistical models. |
| OR-Tools | Turn forecasts into staff schedules, vehicle routes and capacity plans under real constraints. |
| SimPy | Simulate queues and workflows to test how a new plan performs before using it. |
Public data to practise on
- NASA C-MAPSS Jet Engine Simulated Data: Simulated engine sensor runs to failure; predict remaining operating cycles to prioritise maintenance.
- UCI Bike Sharing: Hourly and daily bike rentals with weather and calendar features; forecast demand to plan fleet capacity.
Your first project
Download the NASA C-MAPSS engine data, predict remaining useful life with gradient-boosted trees on rolling sensor averages, and compare it with predicting the training-set mean using RMSE.
Common questions
What is the difference between operations research and machine learning?
Operations research finds the best decision under constraints, for example with optimisation or simulation. Machine learning estimates the uncertain inputs, such as demand or travel time. Many operational systems use both, with predictions feeding an optimiser.
What is predictive maintenance?
Predictive maintenance uses sensor and maintenance records to estimate when equipment is likely to fail, so it can be serviced before a breakdown rather than on a fixed calendar. It needs failure history or run-to-failure data to learn from.
Do operations roles need programming?
Many planning roles still run on spreadsheets, but modelling work increasingly uses Python or R. Even basic Python lets you rebuild a forecast, test it on past data and automate the refresh.
Courses for these skills
- Practical Machine Learning in PythonUnderstand anomaly scores and model assumptions.51 published lessons · more lessons in production
- Maths for Machine LearningConnect optimisation objectives and constraints to a prediction-driven decision.26 published lessons · all planned lessons open
- Data Tools: NumPy, pandas, Matplotlib and scikit-learnAlign process signals and operating metadata.23 published lessons · all planned lessons open
New to programming or maths? Start with Python, or follow the machine learning path in order. Published lessons are free to read; compare Free and Pro.