Service overview
Machine learning, predictive maintenance, computer vision, sensor analytics and forecasting. This service is part of that division and is scoped individually around what you need to learn, decide or publish.
Forecasts feed decisions: how much power will the grid need tomorrow, how much will a wind farm generate this week, when will a component wear to its limit? Machine learning can capture nonlinear relationships between forecasts and their drivers, including weather, calendar and operating variables, and in some problems it outperforms classical statistical models. In others it does not, and the only way to know is to compare them fairly.
Fair comparison is the core of this service. Models are evaluated out-of-sample, over several forecast origins and conditions, against simple baselines such as persistence and seasonal naive forecasts, and against classical time-series models. The forecasts are probabilistic where the decision needs it, because the range of possible outcomes is often more important than the central estimate.
If you plan to publish the outcome, see manuscript & publication support, for example computational & simulation manuscripts or original research article writing.
What this service covers
Problems this service can address
- Forecast skill that disappears out of sample
- Exogenous factors such as weather that must be included
- Need for prediction intervals, not just point forecasts
Methods and analytical approaches
- Feature design with lagged and exogenous variables
- Tree-based and neural forecasting models
- Probabilistic and interval forecasting
- Rolling-origin evaluation against baselines
Applications
- Renewable generation forecasting
- Demand and consumption forecasting
- Degradation trend forecasting
Relevant engineering disciplines
Scope, inputs and deliverables
Scope of work
- Data audit and definition of the prediction or decision task
- Model development with structure-aware validation and baselines
- Evaluation, error analysis and uncertainty reporting
- Documentation sufficient for others to reproduce the work
- Feasibility check on: historical data with timestamps
Typical inputs required from you
- Historical data with timestamps
- Exogenous data sources
- Forecast horizon and decision use
Deliverables
- Forecasting model and baseline comparison
- Interval estimates and error metrics
- Documentation of data splits
Highly specialised requests are first checked for feasibility from this information. Please do not send confidential data through the public form; describe it first and agree secure transfer afterwards.
Project workflow and method
- 01Enquiry and requirement assessment
- 02Technical scoping and proposal
- 03Execution
- 04Quality review and delivery, with an agreed round of revisions
Read the full project workflow.
In depth: how the work is done
Framing the forecast problem
We define the target, the horizon, the update frequency and the decision that will use the forecast. A day-ahead forecast and a five-minute-ahead forecast are different problems with different drivers and information sets. We determine what information will really be available at the time of forecasting, such as weather forecasts rather than weather observations, because using future information that would not be available leads to invalidly good results.
Feature design with lags and exogenous variables
Features include lagged values of the target, rolling statistics, calendar variables such as hour, weekday and holidays, and exogenous drivers such as temperature, irradiance or production plans. Physical knowledge guides the choice, for example clear-sky irradiance for solar forecasting. Features are built in a way that strictly respects time: a value at time t must only use information known at t. Documentation of each feature supports review and reproduction.
Model families
Gradient-boosted trees and random forests handle mixed features and nonlinear effects well on tabular data. Neural networks, including recurrent and attention-based architectures, can model complex sequences but need more data and tuning. Classical models such as exponential smoothing and ARIMA remain strong baselines, particularly for short series. We compare approaches and favour the simplest that performs adequately, because simpler models are easier to maintain and to explain.
Probabilistic and interval forecasts
Point forecasts hide uncertainty. Quantile regression, conformal prediction, ensembles and parametric distributions provide prediction intervals or full predictive distributions. These are evaluated for calibration, whether the stated probability matches the observed frequency, and sharpness, the narrowness of the intervals. Decision-makers use them to set reserves, safety stocks or maintenance margins in ways a point forecast cannot support.
Rolling-origin evaluation against baselines
Evaluation repeatedly forecasts from successive origins, retraining as in operation, and aggregates errors across the whole period and across conditions such as seasons and extreme events. Metrics are chosen to match the decision, such as absolute error, scaled error or pinball loss for quantiles. Results are reported against baselines with confidence intervals on the difference, so that apparent improvements are not due to chance. Performance in rare and extreme conditions is shown separately.
Quality control and good practice
Quality-control approach
- Data are audited for quality, leakage and drift before modelling.
- Validation respects data structure (by specimen, machine or time) and includes simple baselines.
- Performance is reported with uncertainty and failure cases, not only headline scores.
- Model limits and the conditions under which it should not be trusted are documented.
Common pitfalls this service helps avoid
- Using future information in features. Inputs unavailable at forecast time make validation results unrealistically good.
- No baseline comparison. Without persistence or seasonal baselines, the value of a complex model cannot be judged.
- Evaluating only on average error. Large errors in rare conditions may matter most to the decision.
- Ignoring non-stationarity. Changes in equipment, behaviour or climate can degrade a model that was valid historically.
What good reporting looks like
A forecasting report states the target and horizon, data and features with their availability at forecast time, candidate models and baselines, the training and evaluation scheme, accuracy results with confidence and by condition, probabilistic calibration and sharpness where applicable, example forecasts and failure cases, and limitations including non-stationarity and extrapolation. Code and settings are documented. The report does not guarantee future accuracy and recommends monitoring and retraining procedures.
An illustrative scenario
Hypothetical example written for explanation. It is not a client case, a result or a testimonial.
Imagine a researcher forecasting day-ahead solar generation for a small plant. Gradient-boosted trees using lagged generation, clear-sky irradiance and forecast cloud cover are compared with a persistence baseline and a seasonal model. The tree model reduces average error noticeably relative to persistence, though its advantage disappears on days with rapidly changing cloud. Quantile outputs give prediction intervals that cover the observed values close to the nominal frequency. The report presents errors by weather type and notes the limited two-year history.
Questions to settle before you begin
- What decision does the forecast support, and at what horizon?
- Which inputs will really be available when the forecast is made?
- What baselines must the model beat?
- Do I need intervals as well as point forecasts?
- How will I monitor performance after deployment?
Key terms
- Persistence forecast
- A baseline that predicts the next value to equal the latest observed one.
- Exogenous variable
- An external input that influences the forecast target.
- Pinball loss
- A metric for evaluating quantile forecasts.
- Calibration
- Agreement between predicted probabilities and observed frequencies.
- Forecast origin
- The time at which a forecast is issued.
Working together on AI, Machine Learning & Data Science projects
AI and data science engagements start with the decision the model is meant to support and with an audit of the data: sensors, sampling, labels, missing values, the structure of the data by specimen, machine or time, and how future data will differ from past data. We ask for a representative sample and a description of the process. Validation is designed around that structure to avoid leakage, and results are compared with simple baselines. Reports give performance with uncertainty, failure cases and the conditions under which the model should not be trusted. No performance figures are promised in advance.
Frequently asked questions
Why compare against simple baselines?
Many apparent gains vanish against a seasonal or persistence baseline. Showing the comparison keeps claims honest.
How far ahead can you forecast?
The reliable horizon depends on the physics and data. Accuracy typically declines with horizon, and the evaluation reports it as a function of lead time.
Do I need weather data?
For weather-dependent quantities, yes. The quality and availability of forecasts at prediction time should be documented.
Can you forecast with limited history?
With short histories, simpler models and transfer approaches are more appropriate, and the uncertainty is larger.
Will a neural network be best?
Not necessarily. The comparison against simpler models will show what is justified.
How is this service scoped and priced?
After your enquiry we review what you need, ask for missing information and send a written scope and quotation. Price depends on complexity, deliverables and timeline. See the project workflow.