Definition
Load Forecasting is based on historical electricity usage, weather, holidays, and special events, using statistical or machine learning models to estimate electricity load for future periods/areas, supporting scheduling, orderly charging, and demand response.
According to the scale, it can be divided into ultra-short-term, short-term, medium-term, and long-term; new loads such as electric vehicles and distributed photovoltaics have increased dependence on real-time data collection and model updates.
Technical Features
Prediction quality depends on data completeness and timeliness. Smart meters, collectors, and EV charging stations transmit load curves back via the Internet of Things for training and rolling correction; errors directly affect the economic efficiency and safety margin of scheduling.
Use Cases
Used for grid dispatch, distribution network planning, orderly charging, energy storage coordination, demand response, and electricity markets, etc. SimtoLink provides stable cellular connections for acquisition sides such as AMI and pile ends, ensuring reliable prediction inputs.
Key points for cooperation with IoT SIM cards
- The reporting cycle matches the concurrency of meter reading/pile side with the data plan level, avoiding channel congestion before billing periods or scheduling windows.
- Private networks or dedicated APNs isolate the measurement side from public network services, facilitating whitelists and audits.
- ICCID is bound with meters, piles, and transformer area ledgers, supporting abnormal offline location tracking and reconciliation.
- In addition to "online," sampling acceptance is performed for DNS/TLS/long connections to avoid a situation where dialing succeeds but services are unavailable.
A Missing Sample Is Not Zero Load
What a forecast input means for the collection SIM
Load forecasting uses an IoT SIM to deliver load curves from meters, collectors, and charging posts into the model on time. The technical meaning is a power or energy series with time. A missing point should be recorded as missing, not as zero. Zero is read by the model as no consumption in that interval, and both dispatch and orderly charging then build the next plan on a bias. The SIM does not run the forecast. It only affects whether the input is on time, complete, and still carrying the original meter id. Private or directed access is covered under VPDN, and it is used to separate measurements from unrelated public traffic.
Ultra-short-term forecasts are more sensitive to lateness. If a late point can still catch the next calculation cycle, insert it at the original time. If it cannot, leave it missing and let the model apply its own rule. The communication layer must not fill numbers on its own. A longer-term forecast may accept a sparser sample. That does not make missing samples on a short-interval meter acceptable loss.
Conditions for meter and charger backhaul
Report times are staggered away from congestion on the hour, the SIM is bound to the meter or the charger, and after a disconnect the backfill uses the original time. If a large number of terminals all dial at minute zero of the hour, the cell and the platform congest together and the curve shows a sheet of gaps. Size the plan by the number of curve points, not by "hours online." Staying online for a long time without sending a curve does nothing for the forecast. Charger power moves quickly. The sample interval should be written at activation. It must not use the same default period as a residential meter.
Feeder, line, and charger-group ids have to travel in the message. If the SIM changes and the meter number does not, the new curve becomes a series with no history, and the model treats it as new load. For roaming or cross-region chargers, confirm the attach policy, so a vehicle at a service-area boundary does not lose the same stretch of curve every day.
A run of missing samples caused by congestion on the hour
A common failure fills a gap with the average of the points on either side, or with zero, and then sends the filled value into training. The model learns a shape of consumption that was never there. Another is a backfill that cannot cover: when the late real point arrives, the zero already filled will not give way, and the real peak disappears. Watching only that the SIM is online, and not the point count, still shows the device healthy after days of continuous gaps.
Leaving a debug meter's dense reports on a production SIM crowds the top-of-hour window of other meters on the same cell. Acceptance should create one hour of missing samples: the curve shows a gap rather than zero, the late point returns to its original time, and a second series with the same meter number is not created.
Before a curve enters the model it also has to be checked for multiplier, number of phases, and freeze time. None of that is in the SIM's online state. If a meter reports a secondary value and the platform reads it as a primary value, the load is scaled up or down, the forecast follows, and communication has fully succeeded. The message should carry the multiplier, or send the primary value directly, and that should be fixed in the ledger. When the meter is replaced, the multiplier changes with the meter number, not with the SIM number. If one phase of a three-phase meter is missing and the missing phase is filled with zero, total power is short by that phase. Mark the missing phase on its own. Whether total power is computed from the remaining phases or the whole interval is void has to be written in the rules beforehand. The backfill program must not decide it. Do not mix frozen energy and instantaneous power reported after time sync on one curve. The first is energy over an interval. The second is an instant. Both use time on the horizontal axis, and they do not mean the same thing. Under congestion on the hour, frozen values may go first and the instantaneous curve may follow later, but the type field on both must not be omitted. The SIM only guarantees that the bytes arrive. A wrong type trains energy as if it were power.