The full power of TDengine, now free forever for up to 5,000 tags.

Explore More

How Can Engineers Forecast Industrial Time-Series Data?

Juno Qiu

November 10, 2026 /

Industrial time-series forecasting uses historical measurements to estimate how a process variable may change over a future period. A process engineer can use a forecast to prepare for a decision, provided the forecast has been evaluated for the right time horizon and operating conditions.

Start with the decision, not the model. If an engineer needs to plan treatment capacity for the next shift, a forecast of influent flow over that period may be useful. A forecast for a different interval, or one that performs poorly during the conditions expected on the shift, may offer little help.

Define the measurement and the decision

Consider a hypothetical water treatment facility. Before the next shift, a process engineer wants to review expected influent flow and discuss whether the available treatment equipment is sufficient. The forecast would inform that discussion; operators would still follow the site’s approved operating procedures.

The first questions are practical:

  • Which flow measurement represents the incoming load at the relevant point?
  • Does the team need hourly values through the shift, a shift total, or another measure?
  • How far ahead must the forecast extend for the team to prepare capacity?
  • How much error can the decision tolerate before the forecast becomes misleading?

The answer determines the forecast horizon, the period into the future for which values are estimated. A forecast that performs well one hour ahead may not be reliable for a full shift. Evaluate it at the horizon the decision actually needs.

Forecasting influent flow is also different from predicting equipment failure. A forecast may estimate a future process measurement; it does not by itself establish whether a pump or another asset will fail.

Check whether the historical data supports the task

Historical values need enough context to be interpreted. For influent flow, confirm the measurement location, units, sampling interval and time zone. Look for gaps, delayed readings, instrument changes and periods when the recorded signal may not represent normal operation.

Then examine the conditions behind the data. Flow may follow daily or weekly patterns, but a scheduled shutdown, a change to the collection system or an unusual weather event can alter the relationship between past and future values. If a period is excluded from evaluation, record why. Excluding every difficult period can make a forecast look better than it will be in use.

Some forecasting methods can use additional variables. Any input used for a future forecast must be available when the forecast is made. For example, a weather forecast available at the time may be considered, but the actual rainfall observed later must not be substituted into a historical test. That would give the test information the engineer would not have had when making the decision.

Preparing historian data for industrial AI models covers timestamps, data quality and operating context in more detail. The data needed for a particular forecasting method will depend on the method and the decision.

Compare the forecast with a simple baseline

A forecast should provide useful information beyond an accessible reference. For a stable measurement, the latest observed value might be a starting baseline. For a signal with a consistent daily pattern, the value from a comparable prior period may be more appropriate.

A baseline is not automatically correct. A previous day may have had different weather or operating conditions. Its purpose is to give the team a transparent comparison: does the forecasting method improve the decision enough to justify using it, including when conditions are difficult?

Test using past forecasts, not future knowledge

Chronological backtesting evaluates what a method would have predicted at earlier points in time. Select a past forecast origin, use only the information that would have been available then, and compare the forecast with the measurements that followed. Repeat this across representative periods and at the horizon needed for the next-shift decision.

For the treatment facility, the team could test several past shift starts and compare each full-shift forecast against the subsequently recorded influent flow. Include different operating conditions where the available record permits. Compare the results with the same baseline over the same periods. The time-series cross-validation method provides a formal version of this rolling-origin approach.

Look beyond one overall error figure. Ask whether errors are consistently too high or too low, whether larger errors occur at certain times, and whether the method handles periods of changing flow. A forecast may appear acceptable on average while missing the conditions most relevant to equipment-capacity planning.

For example, a model might track ordinary shifts reasonably well but repeatedly underestimate influent flow after heavy rain. Its overall average error could look acceptable while the underestimates occur precisely when the team needs to assess capacity. The engineer should examine those periods separately and use the established planning process if the forecast is not reliable enough for that decision.

Backtesting is an engineering evaluation method. The exact way a team performs it depends on its tools and workflow; it should not be assumed to be a one-click feature of every forecasting interface.

Know when not to rely on the forecast

A forecast should receive less weight when the available history has serious gaps, the measurement is suspect, or current operations differ substantially from the periods used for evaluation. It may also be unsuitable if it does not outperform a simple baseline or its errors are too large for the capacity decision.

Uncertainty matters. A single predicted curve does not show every plausible outcome. Review the range of past errors for comparable conditions and consider what the team would do if actual flow differs from the forecast. Historical error is evidence for that discussion, not a guarantee that future errors will stay within the same range.

TDengine IDMP documents time-series forecasting with TDgpt, including forecasts based on historical values. Its LLM connection documentation distinguishes TDgpt forecasting from features that use an external language model. Model options and configuration depend on the deployed version. The forecast still needs to be assessed against the facility’s own data and decision requirements.

For the next shift, the engineer can share the forecast, its comparison with the baseline, and the known limitations with the operations team. If the evidence is insufficient, the team should use the established planning process rather than treating a forecast as a substitute for operational judgment.

Related Articles