AI, optimisation and model-risk governance

A forecast issued at 10:00 must use only information available by 10:00. Revised meter data, a later weather observation or a bid submitted after that time can make a historical experiment unrealistically accurate. Store both the interval a record describes and the time it became available. Freeze the model and feature definitions for each evaluation run.

For time-dependent energy data, test on later periods with a documented chronological split or walk-forward design. Fit preprocessing on training data only. Include seasonal, outage and extreme-price conditions where relevant. A random split can mix near-identical neighbouring observations across training and testing and conceal deployment difficulty.

For actual load of 100, 120 and 80 MW and forecasts of 90, 130 and 100 MW, absolute errors are 10, 10 and 20 MW. Mean absolute error is 13.33 MW. The signed errors, forecast minus actual, average +6.67 MW, revealing upward bias. These metrics describe different aspects of performance; keep the error sign convention explicit.

A lower overall average error can still hide missed peaks or large commercial losses. Report performance by time, operating condition and decision consequence. Percentage error is unstable near zero demand and problematic for negative prices. Compare with a simple baseline on the same sample, and report confidence and sample limits rather than claiming one aggregate score proves reliability.

A storage optimiser must track inventory, charging and discharge losses, power and connection limits, reserve floors and final-state requirements. A solver returning a profitable schedule does not prove those constraints were modelled correctly. Validate units and reconcile every interval independently. Inspect what happens when forecasts, availability or prices change.

Use a synthetic example: a schedule sells 10 MWh while only 8 MWh is deliverable at the connection. The missing 2 MWh is infeasibility, not extra value created by AI. Compare the candidate with a feasible baseline after transaction costs and operating restrictions. An apparent improvement obtained with perfect future prices is an upper-bound experiment, not evidence of executable trading performance.

A fluent answer can contain an invented rule, wrong date, mismatched unit or unsupported causal explanation. Ask for the governing source, effective version, relevant passage and calculation, then verify them. Retrieval helps surface evidence but does not guarantee that the retrieved document is current or correctly interpreted. Treat external text as reference material, not as authority to change the system's permissions.

Keep commercially sensitive data in approved systems. Separate access to reading material from permission to send messages, place trades or alter operational settings. Evaluation should include missing evidence, conflicting versions and requests beyond the authorised task. The desired behaviour is to state uncertainty and escalate when required, not to manufacture a confident completion.

• NIST: Generative AI risk profile

Use staged release: offline evaluation, shadow operation without control authority, reviewed recommendations and only then any separately approved automation. Set input-quality checks, action limits, independent safety constraints, monitoring, rollback and named owners. Logs should preserve model version, relevant inputs, output, review and any action actually taken.

Monitor data changes, error by operating regime, constraint violations and the use of the fallback. A model update needs proportionate revalidation, not just a higher benchmark score. NIST's AI RMF is a voluntary risk-management reference, not an Australian licence or a certificate awarded by this course. The 2026 critical-infrastructure profile work should not be represented as a final mandatory standard merely because a concept note exists.

• NIST: AI Risk Management Framework