At a commercial campus site in north Bengaluru, a queue begins at 5:20 PM. Chargers do not hit 100% utilization until 6:10 PM, yet drivers unable to find a spot are already parking on the street and contacting facilities management. The dashboard shows no issue until 6:10, when all 18 connectors enter Charging simultaneously. The queue has then grown for 50 minutes.
Demand forecasting closes that gap. Saturation arrives at 6:10, but session initiation patterns reveal it three to four hours earlier. With three-day-ahead forecasting, Monday data can flag a Wednesday peak, leaving time to notify drivers, adjust pricing, or pre-position a mobile charging unit if available.
Why Current Utilization Misleads
Operators commonly track charger utilization in real time: the share of connectors in Charging. This helps detect faults, such as an unexpected drop to zero, but offers little for capacity planning. Once utilization is high, the decision window has passed.
Capacity decisions need session initiation rate across time windows, not current utilization. A site moving from 3 starts in the 4-5 PM hour to 7 in the 5-6 PM hour and 14 in the 6-7 PM hour is heading toward saturation, even with utilization at just 45% at 5:30 PM. Initiation rate leads; utilization percentage follows.
Notifications, overflow routing, and pricing changes must occur before saturation. Utilization-based action is reactive, while session initiation trends create time to respond.
What Session History Shows
A single site's session data exposes three stable patterns once 6 to 8 weeks of history are available.
First is the day-of-week signature. Office campuses typically record 20 to 30% lower utilization on Mondays and Fridays than Tuesday through Thursday. Commercial real estate sites also differ sharply between weekends and weekdays. This consistency lets a simple day-of-week model explain much of session-volume variance.
Second is time-of-day distribution. Commercial and office sites usually cluster session starts in a morning arrival band, 7:30 to 10:30 AM, and an evening peak, 4:30 to 8:30 PM. Their sizes vary by site type, but the timing is stable enough to forecast. A 3-hour moving average for the same time-of-day and day-of-week window provides a practical baseline.
Third is occupancy duration. With a historical average session of 2.8 hours and 16 connectors, theoretical saturation begins when more than 16 sessions start within the prior 2.8-hour rolling window. Duration data helps estimate queue onset, rather than merely flagging "high utilization."
Building a Three-Day Forecast
A useful site forecast needs no machine learning model. It needs historical session data and a query that estimates hourly volume for a target date, adjusted for day-of-week and known calendar effects, including local holidays and nearby major events.
Our pilot network forecast takes a rolling 4-week average of session starts for each (day-of-week, hour) bucket. For Wednesday 6 PM, it averages the last four Wednesday 6 PM windows, then applies a trend adjustment when utilization has grown over those 4 weeks. It forecasts each hour across the next 3 days and compares demand with charger capacity to generate a congestion probability signal.
A congestion probability above 60% for any hour sends an alert 72 hours ahead. Operators can notify drivers through their app, such as "high demand expected Thursday evening, consider charging before 5 PM", and change pricing if dynamic rates are supported. In our pilot network, this warning process cut peak-period queue complaints by roughly 35% versus the same sites without proactive notifications in the prior quarter.
Limits of Three-Day Forecasting
We are not dismissing three-day forecasts. They suit operational changes, pricing nudges, and driver communications. This article concerns that operator profile. Permanent capacity still needs corroborating evidence, not a forecast alone.
Pattern-based forecasts can fail on important events. Local holidays in India vary across the calendar, while state holidays such as Karnataka Rajyotsava in November may be absent from generic public holiday APIs. A site usually at 80% utilization on Tuesday may reach 30% on a state holiday unless you maintain a local holiday calendar.
Nearby concerts, sporting events, and large malls' weekend events can create spikes invisible to a session-history-only model. In areas with heavy event foot traffic, omitting an event signal will consistently underforecast event days.
A growing registered driver base also weakens pattern stability. Adding 40 drivers in one month, common after a nearby residential complex onboards, can make session history from 8 weeks ago unrepresentative of next week's demand. Weight recent history more heavily and widen uncertainty bounds during rapid growth.
Capacity Decisions: Forecast Plus Evidence
Add charger capacity based on a persistent pattern, not one forecast: high congestion probabilities across a 3 to 4-week period, queue events in session data, including starts more than 15 minutes after driver app initiation, and direct driver feedback. Any one signal can mislead; together they support a defensible capital decision.
The forecast is the earliest warning; session patterns and driver feedback confirm it. Acting on forecasts alone can add capacity too soon, while waiting for complaints delays action. The deciding variable is how early confirmed congestion arrives.