Solutions · Energy and industry

Forecasts that say when they are guessing.

Load, output and failure models work until the conditions change. Kaimei shows what each forecast leaned on, in its own units, and flags the hour the inputs leave the range the model was trained in.

Data
8,568 hoursof city load and weather, from 52,416 ten-minute readings
Accuracy
2.8%mean absolute percentage error on held-out hours: 1.9 MW
When July arrives
1.9 → 4.2 MWmean absolute error more than doubles outside the training season
Why
PSI 6.77on temperature: the inputs left the range the model learned in

Where it fits

For forecasts, sensors and inspection lines.

Grid

Load and generation forecasting

What a forecast leaned on, hour by hour, in megawatts: yesterday's load, the hour, the weather.

Assets

Predictive maintenance

Why this unit was flagged, from its sensor features, and which sensor has drifted since the model was fitted.

Quality

Visual inspection

Grad-CAM and occlusion on a defect classifier, with a check that the model looked at the part and not the background.

Process

Yield and process models

Behaviour curves show how predicted output responds to each setting, and whether that response is monotonic.

Field

Sensors and units

A sensor that starts reporting in another unit shows up as input drift within one batch.

Sites

Isolated environments

Run Kaimei on your own infrastructure, or upload shapes only and keep the weights on site.

A worked example

A day-ahead load forecast, opened up.

Public dataPower Consumption of Tetouan City, UCI Machine Learning Repository (Salam & El Hibaoui, 2018).load-forecastscikit-learn · neural network

A small neural network, two hidden layers behind a scaler, forecasts hourly city load in megawatts from eight inputs: hour, weekend, four weather readings and the load a day and a week earlier. It is trained on January to June 2017, 4,176 hours, and then meets July.

load-forecast · Explain
scikit-learn · neural network
raiseslowers
Power Consumption of Tetouan City POST https://api.kaimei.ai/v1/explanations

One forecast

A good hour and a bad one, in megawatts.

At 21:00 on 30 June the model forecasts 105.5 MW against an actual 104.4. The explanation starts from the 70.6 MW average and adds +20.3 for yesterday's load and +12.9 for last week's.

At 11:00 on 24 July it forecasts 87.2 MW and the city draws 104.0. The explanation shows why: it is 29.4 °C, and temperature adds only +0.7 MW. The model leans on a lag of 88.6 MW (+9.5) because nothing in its training months taught it what that heat does to demand.

Mean absolute error · load-forecastMW
1.9held out
4.7July wk 1
4.1July wk 2
3.9July wk 3
4.3July wk 4
2.8% error on 836 held-out hours from January to June; 4.8% across July, when the city averaged 27.2 °C and 88 MW. The model under-forecasts July by 2.0 MW on average.

Out of range

The inputs moved before the error did.

Kaimei compares the latest 500 logged hours with the reference hours, one input at a time. By late June temperature has already drifted from a reference that starts in January. Through July, wind speed and both lagged loads follow. Nobody needs to wait for the actuals to learn that the forecast is extrapolating: the alert arrives with the inputs.

Late Junesignificant
Temperature3.67
Wind speed0.98
Load, a week earlier0.35
Solar irradiance0.33
Load, 24 hours earlier0.31
Humidity0.13
Hour0.00
Day0.00
Recent traffic has significant drift (PSI 3.67). Shifted inputs: temperature, wind_speed, load_168h_ago, solar_irradiance. 500 logged predictions against 4,000 reference rows.
Julysignificant
Temperature6.77
Wind speed3.55
Load, 24 hours earlier2.03
Load, a week earlier1.40
Humidity0.80
Solar irradiance0.71
Hour0.00
Day0.00
Recent traffic has significant drift (PSI 6.77). Shifted inputs: temperature, wind_speed, load_24h_ago, load_168h_ago. 500 logged predictions against 4,000 reference rows.

Population stability index per input. Bars are cut off at 0.5; the line marks 0.25, where Kaimei calls drift significant.

What it learned

Mostly yesterday, a little weather.

load_24h_ago is the most influential input, followed by load_168h_ago and hour. Together they carry 84% of the measured importance (permutation importance over 2000 rows).

Temperature holds 3.3% of measured importance. For a model that will face a summer, that is the finding: retrain with a full year, or give the forecast a temperature-driven term.

  • Higher load_24h_ago increases load_mw, most sharply between 80.19 MW and 91.88 MW (range 58.4 → 86.5).
  • temperature has a non-monotonic effect on load_mw: it bottoms out around 14.39 °C (range 68.3 → 70.1).
  • hour has a non-monotonic effect on load_mw: it bottoms out around 4 (range 67.2 → 72.1).
Global importance · load-forecastpermutation · 2,000 rows
Load, 24 hours earlier43.7%
Load, a week earlier31.3%
Hour8.9%
Solar irradiance4.7%
Day3.6%
Temperature3.3%
Wind speed2.5%
Humidity1.9%
load_24h_ago is the most influential input, followed by load_168h_ago and hour. Together they carry 84% of the measured importance (permutation importance over 2000 rows).

Inside the network

Every layer and weight, drawn.

An uploaded network is drawn layer by layer, and a trace walks one prediction through it: the inputs, each weight and bias, each activation, the output. Repeated blocks fold, so a large model stays readable.

  • Keras, PyTorch and scikit-learn networks, with scalers folded in
  • client.models.activations returns what each layer saw
  • Shapes-only uploads are still drawn; the weights stay on site
Model anatomy
Robustness · load-forecast300 rows · 5%-σ noise

Outputs move by 3.8% of their spread under 5%-σ input noise. Most sensitive inputs: wind_speed, load_24h_ago.

Predictions range from 38.9 to 106 (median 70) across 2,000 reference rows.

Behaviour curves for the 6 most important features. 3 respond monotonically; hour has a non-monotonic effect.

Who it is for

Forecasters, operators and engineers.

01

Forecasting and data science

See what each forecast leaned on and how the model responds to every input, and compare next season's retrain with this one.

02

Operations and control rooms

An alert when the inputs leave the range a model was trained in, with the input named, before the error shows up in the actuals.

03

Reliability and quality engineers

Why a unit or a part was flagged, in sensor terms, and whether an inspection model looked where it should.

Coming soon

Next on the roadmap.

What we are building now, and what to use until each one ships.

  • Coming soonSequence and time-series models

    Forecasters that read a history step by step, with each forecast attributed to time steps as well as inputs. Until then, a model on lags, rolling averages and calendar features works.

  • Coming soonMulti-horizon forecasts

    Several outputs per model, so every horizon of one forecaster is explained in a single call. Until then, register one model per horizon.

  • Coming soonDetection and segmentation

    Explanations for the boxes and masks a model draws on inspection imagery, not only for its label. Image classification works today.

  • Coming soonScheduled checks and alert routing

    Drift and performance checks run by Kaimei on a schedule, with alerts sent to email, Slack, PagerDuty or a webhook. Until then, run them from your own scheduler.

See it on your own model.

Bring a forecast or a sensor model. See what it leans on before the season turns.