Grid load forecasting copilot.

Grid dispatch run on static spreadsheet forecasts with no confidence ranges and no audit trail.

Client: An energy operator
Sector: Energy, grid operations

P10 / P50 / P90probabilistic forecasts

Growth goal

Give grid operators forecasts they can dispatch on, with the uncertainty visible.

Operating constraint

Dispatch ran on a single-point spreadsheet forecast: no ranges, no drift monitoring, no record of why decisions were made.

Validated result

P10, P50 and P90 forecast ranges, drift checks, a decision record and an approval gate, demonstrated in the tested workflow.

The full story

Baseline: Static spreadsheet forecasts, one number, no audit trail.
System: Probabilistic forecasting with P10, P50 and P90 ranges, drift checks, decision logging and an approval-gated copilot.
Controls: Nothing dispatches without approval. Drift is monitored and every decision is logged with its forecast.
Result: Ranges, monitoring, decision record and approval gate shown working in the tested workflow.

In depth

The operator had to plan dispatch from thin forecasts

The operator’s grid was carrying more variable load, more distributed generation and more commercial demand. Dispatch teams had to decide what to bring online, what to hold in reserve and where to keep headroom. They were accountable for keeping the grid stable and the economics sensible. On paper, the system could carry the growth. In the control room, operators were still looking at a single spreadsheet number and making judgement calls with no way to see how uncertain that number was. Growth was adding pressure without adding visibility.

Dispatch decisions rested on a static, single-point load forecast prepared in advance. There was no confidence range around that value. There was no indicator of how recent behaviour compared with the forecast that produced it. Once an operator chose a dispatch plan, there was no structured way to record what they had seen and why they acted. The process relied on individual experience and memory. The operator wanted to grow with more intermittent supply and more complex demand, but the forecasting and decision record had not caught up.

Single point forecasts and no audit trail hit their limit

The baseline was clear. A spreadsheet held the latest forecast as one number per interval. Operators read that number, applied their own mental buffers, then chose which assets to dispatch. If conditions changed, they adjusted on the fly. There was no formal view of upside or downside risk, no warning if the forecasting model started to drift and no central log of how decisions matched the forecast seen at the time. The organisation could not easily review past days and understand whether it had been taking more risk than intended.

This set a hard limit on how much new variability the grid could handle with confidence. As more variable generation and flexible demand came on, planners wanted to know not just the best guess, but the plausible low and high cases. Compliance and operations leaders wanted a clear audit trail that showed which forecast an operator saw, what action they took and what approval it carried. Without that, it was hard to tune policies, hard to defend decisions and hard to change the playbook for new market conditions. The operating model was carrying too much silent risk.

SIEL built probabilistic forecasts into a guided workflow

Working with the operator, SIEL focused on the core workflow that turned forecasts into dispatch decisions. The goal was not to replace the dispatch process, but to give it better inputs and better records. SIEL built a probabilistic forecasting system that produced P10, P50 and P90 ranges for the relevant grid intervals. These ranges gave operators a clear view of the baseline scenario and the plausible lower and higher load outcomes that mattered for risk. The forecasts were wired into a copilot interface that sat alongside existing tools, not in their place.

Inside that copilot, each forecast came with its P10, P50 and P90 values and checks that watched for drift in the underlying model behaviour. When an operator prepared a dispatch plan, the copilot surfaced the forecast ranges and prompted them to review how their plan sat against those scenarios. The operator could adjust their plan with that context in view. Once they were satisfied, the copilot moved the plan into an approval-gated step. The system itself did not dispatch. It prepared a complete, explainable proposal and held it for the right person to approve. Every step wrote to a decision log tied to the underlying forecast.

Approvals, drift checks and decision logs kept people in charge

Control was central to the design. Nothing in the POC dispatched without explicit human approval. The copilot collected the proposed dispatch action, the P10, P50 and P90 forecast values that informed it and any notes the operator chose to add. It then presented the package for approval. Approval was a distinct, recorded step. This created a clear point where responsibility was taken and where oversight could focus.

To manage model risk, SIEL added drift checks around the forecasting system. These checks watched how forecasts and realised behaviour compared over time within the tested workflow. If they started to diverge beyond agreed bounds, that was visible. This gave the operator a way to treat the model as a monitored component, not a black box. Decision logging meant that, later, anyone could see the forecast that was available at the time, the action chosen, who approved it and how that related to the uncertainty range. The operation stayed firmly human-led, with the system providing structured support and a reliable record.

Forecasts became something operators could build on

The proof of concept showed that probabilistic forecasting and a copilot workflow could fit the operator’s world. In the tested workflow, operators could now see P10, P50 and P90 forecast ranges rather than a single guess. Drift checks ran around the model so its behaviour stayed visible. Every dispatch decision in that flow wrote a structured record tied to the forecast it used. Nothing moved to execution without an approval gate being passed. The operator had a clearer, auditable bridge between forecasts and dispatch, without losing human judgement.

For a company in a similar position, the pattern is repeatable. You keep your operators in charge of dispatch. You add a forecasting system that shows uncertainty in concrete numbers, not just as a feeling. You wrap that system in a workflow that logs what people see and what they do, with approvals where they matter. SIEL’s role is to connect those pieces so they work with the systems you already run and the constraints you already live with. Six months on, the aim is simple. Your team is still making the calls. They are just making them with clearer forecasts, visible model behaviour and an audit trail you can trust.

Bring us a workflow like this

Fixed fee, agreed before we start. A senior engineer replies within two business days.