Alerts and drift

Being told about a consumption drift, without drowning in false alerts

The problem with an alerting system is not detecting: it is not crying out for nothing. An alert that fires every night ends up filtered by a mail rule, and the monitoring dies with it. Softee treats noise as a first-order setting, not an advanced option.

For energy managers and operations teams who want to learn about a drift within days, rather than from next quarter’s invoice.

Request a demonstration

What can an alert be set on?

On four families: consumption drift, the invoice, the period and the sensor itself. The last two matter as much as the first: zero consumption on a working day and a silent sensor are information, and often the most urgent kind.

An alert’s scope targets sites, indicators, sensor types or specific sensors. Its granularity runs from the minute to the month, its aggregation from the hour to the month, and a time window can restrict it — enough to watch consumption outside occupancy hours without being woken by the normal consumption of the working day.

What does an alert compare the observed value against?

Five comparison modes are available: a threshold, a sum, an average, a drift or a data spike. The threshold is set in absolute value or as a percentage, across three severity levels: information, warning, critical.

Four time references are possible: the previous month, the same month last year, the previous week or the same week last year. The two year-on-year references are the ones that count in energy, since they compare a period against its seasonal equivalent rather than against the period before it.

How does Softee avoid false alerts?

Through two adjustable mechanisms, distinct and combinable. Neither is machine learning: they are settings you can read and can disagree with.

Quiet period

After firing, the alert goes silent for a duration you set. The default is 24 hours.

Deferred check

A second check can be required, some time after the first, before the alert genuinely fires. An isolated spike wakes nobody; a settled drift does.

Minimum severity

The severity level at which the email goes out is set per alert. An information alert can feed the log and nothing else.

Who is notified, and how?

Recipients are defined alert by alert, with the minimum severity level that triggers the email. A critical alert can therefore write to the on-call team while an information alert stays in the log. Alert activation and email sending each have their own switch at organisation level.

In parallel, everything lands in an event log: an alerts dashboard, history per alert or per site, and acknowledgement of an event singly or in bulk. Event retention is set in months.

When are alerts evaluated?

Every day, automatically, by a scheduled job that runs after collection and calculation — the order matters, otherwise you would be evaluating yesterday’s data. A user can also force evaluation of one alert or of all of them, which is how you test a rule you have just written without waiting for tomorrow.

What this page does not claim

  • No SMS notification, and no outbound webhook. The alert notification is an email to the defined recipients. The only other outbound path in the product is a message queue, used by scheduled exports.
  • No machine-learning detection. The alert rules are rules: thresholds, sums, averages, drifts, comparisons against a reference. They are explicit and verifiable, and we will not dress them in vocabulary they have not earned.

Frequently asked questions about drift alerts

Four families: drift, invoice, period and sensor. An alert’s scope targets sites, indicators, sensor types or specific sensors; its granularity runs from the minute to the month, its aggregation from the hour to the month, and a time window can restrict it — which allows consumption to be watched outside occupancy hours.
Through two combinable mechanisms: a quiet period after firing, adjustable and set to 24 hours by default, and a deferred second check that can be required before the alert genuinely fires. Thresholds are also set in absolute value or as a percentage across three severity levels, and the minimum level that triggers the email is set alert by alert, which allows notification to be reserved for critical cases.
One of five comparison modes — threshold, sum, average, drift, data spike — and one of four time references: the previous month, the same month last year, the previous week or the same week last year. The two year-on-year references compare a period against its seasonal equivalent, which is the comparison that makes sense in energy.
Every day, by a scheduled job that runs after data collection and the calculation pipeline. A user can also force evaluation of one alert or of all alerts on demand, which allows a rule to be tested immediately after writing it.
No. The alert notification is an email addressed to the recipients defined on the alert, above a minimum severity level. There is no SMS sending and no outbound webhook in the product; the only other outbound path is a message queue, used by scheduled exports.
No, and it is a choice we stand behind rather than a gap we are hiding. The rules are explicit: thresholds in absolute value or as a percentage, sums, averages, drifts, comparisons against a time reference. You can read them, tune them and disagree with them — which is not the case with a model whose reason for crying out is unavailable to you.

Bring us the drift you missed

Consumption that ran away for two months before anyone noticed. We will look together at which rule would have caught it, when, and what your meters’ reporting frequency actually allows.

Request a demonstration