BESS Dispatch: From Forecast Errors to Operating Decisions

A solar-battery dispatch plan decides when stored energy is worth using. Its value depends on what the controller knows at that moment, what it can still change, and which bill or operating constraint it is trying to improve. A simulation that knows tomorrow’s solar production and demand answers a useful design question. An operating controller must make the same decisions before tomorrow has happened.

Forecast accuracy alone does not determine the value of battery dispatch. The cost of an error depends on when it occurs, whether current measurements can change the scheduled action, and whether the battery still has enough power and energy to respond. Research on solar storage makes these connections visible: a modest error at the wrong moment can outweigh many hours of accurate prediction.

What the forecast comparison actually tested

Mirletz and Laws’ 2023 simulation used 2018 Puerto Rico weather and an EnergyPlus hospital load model. The modeled system combined 3.66 MW-ac of PV with a 550 kW, 6,500 kWh battery and a 70% minimum SOC for resilience. [1]

Table I reports NPV reductions of 0.91% for peak shaving, 1.23% for price signals and 8.23% for the labeled 24-hour MPC case when day-ahead forecasts replaced perfect inputs. The full-year case lost 10.7%, but lacked real-time information and replanning. Battery-model and billing differences prevent a direct algorithm ranking. The modeled 25-year battery life was also identified as optimistic. [1]

NPV expresses discounted project value over the modeled lifetime. Its sensitivity to forecast inputs reflects the financial consequences of dispatch decisions, rather than the percentage error in the forecast itself. A small change in annual operating savings can affect that lifetime value repeatedly; the timing of the lost savings also matters because future cash flows are discounted.

The expensive error can be one decision at the wrong time

In Figure 1, the forecast expected strong PV at the 11 a.m. demand peak. Controllers with current information stopped charging; the full-year schedule continued. Although demand charges represented only 9.6% of the original bill, they dominated the savings difference caused by forecast errors. [1]

This is the distinction between average prediction quality and the cost of a decision. An error during an uneventful low-load period may be absorbed without changing the billed maximum. The same error near a billing peak can change the most expensive interval of the month. A forecast metric that averages every interval equally does not express that difference.

For a simple tariff charging a rate r per kW of the highest metered demand in a billing period, the demand component is r × max(Pgrid,t). Here Pgrid,t is the average import over the tariff’s metering interval. The maximum makes the effect uneven: reducing an interval below an already higher maximum may leave this charge unchanged, while creating a new maximum increases it. Actual tariffs can use multiple periods, ratchets or other rules, so the applicable calculation must come from the site’s tariff. Read the billing-input guide.

For solar storage, an expectation of surplus PV and an expectation of a net import peak can imply opposite battery actions. The first can favor charging; the second can favor discharge or at least stopping charge. When the predicted surplus fails to arrive, charging may add to grid import at precisely the time it is most costly. The annual amount of solar energy says little about this coincidence.

Planning and feedback solve different problems

A plan allocates a limited energy budget across future opportunities. Feedback adjusts the immediate action using current measurements. These functions operate on different information: the plan anticipates a later peak, while feedback responds to the gap between the expected and measured conditions now. Good scheduling cannot compensate for a controller that continues an unsuitable action after conditions have changed.

SAM’s documented peak-shaving method builds grid-power targets from load and generation forecasts, then attempts to meet those targets within battery capacity and SOC limits. Its perfect look-ahead option uses the future load and solar resource that coincide with the simulation; custom forecasts separate predicted weather from the weather driving actual simulated production. These options represent different information conditions, even when the equipment specification stays identical. [2]

At a common AC measurement boundary, the power balance is Pgrid = Pload − Ppv − Pbatt, with positive Pbatt denoting discharge and site load including auxiliaries. Losses are accounted for consistently at that boundary. If PV output falls while load stays constant, grid import rises unless battery discharge increases. A fixed battery-power command therefore has a different effect from a grid-import target: one holds the battery action constant, while the other requires that action to change with the site.

Feedback still cannot recover energy already spent. A controller may recognize a late peak correctly yet have insufficient stored energy to reduce it. Conversely, a battery can have ample energy but insufficient discharge power to cover a short, sharp peak. Forecasting determines what the controller expects; scheduling determines how it allocates energy; immediate control determines whether the current action follows the changing conditions.

Why the simplest controller is not automatically the best choice

Mirletz and Guittet’s 2021 study compared dispatch across two modeled California building cases and 625 tariff variations. Price-sensitive dispatch balanced energy charges, demand charges and a cycling penalty; peak shaving and a fixed schedule each performed well under different tariff structures. All cases used perfect day-ahead forecasts and an NMC battery model. The chemistry-specific degradation assumptions limit how the lifetime results transfer to other batteries. [3]

The algorithm also assigned value to energy left at the end of its planning horizon. Its discussion showed that the treatment of accumulated electricity credits and battery replacement could change the preferred strategy. The authors identified further trade-offs between throughput, efficiency losses and demand reduction. [3]

A planning horizon has an artificial end; the operating site does not. If the last stored kWh has no value at midnight in the model, discharging before midnight can appear attractive even when the energy would be more useful the following morning. A high terminal value can produce the opposite tendency, preserving energy despite an opportunity to use it. This explains why two optimizers can choose different actions with identical forecasts: they may assign different value to the state left for the next planning period.

A controller can also track a peak target accurately while pursuing an unsuitable economic objective. A site paying mainly for time-varying energy may benefit from a different energy allocation than a site dominated by demand charges. A protected backup reserve reduces the energy available for either purpose. The interaction between reserve and peak shaving is explored in the solar BESS reserve guide; it changes the feasible dispatch choices before forecast quality enters the calculation.

Why an ideal forecast makes savings look easier

Perfect foresight removes a difficulty that an operating controller must face: uncertainty about which future interval deserves the stored energy. It can preserve energy for a peak whose exact timing is already known. A forecast-driven controller has to act before that uncertainty is resolved. The gap between the two can therefore persist even when both use the same battery and obey the same power and SOC limits.

The timing of information matters as much as its contents. A forecast revised after a cloud front arrives contains knowledge that was unavailable when an earlier charge decision was made. A model using that revised forecast for the earlier decision benefits from hindsight. Similarly, resetting SOC at the start of each day erases the consequences of yesterday’s dispatch. Both assumptions can make a strategy appear more capable than a controller operating continuously.

The reference result is also shaped by the tariff and battery model. A strategy that looks attractive with one treatment of export credits or replacement costs may lose that advantage when those assumptions change, as the 2021 study illustrates. [3] This is why the 2023 forecast comparison is most informative within each algorithm: changing the forecast exposes sensitivity to information, while differences in the underlying models complicate comparisons between algorithms. [1]

Power, energy and control targets impose different limits

The distinction between power and energy is concrete in HUA-X’s PowerFac-261: the liquid-cooled cabinet is rated at 125 kW and 261 kWh, with a 400 V AC connection. [4] Power describes the rate of response; energy describes the stored quantity from which that response draws. A long period of moderate import can consume substantial energy without reaching the power limit, while a brief large peak can exceed the power limit with energy still available.

Control targets add a separate distinction. A battery-power setpoint specifies what the battery should do; a grid-import target specifies an outcome at the site meter. The latter depends on load and PV as well as battery power. If the site meter covers loads outside the controller’s measurement boundary, an apparently correct local response can coexist with an unexpected import peak. The BESS EMS interface guide explains these signal and measurement boundaries.

At container scale, PowerFac-5M provides 5,016 kWh of nominal DC-side battery energy, with PCS and EMS/SCADA forming part of the wider system configuration. [5] That energy figure alone does not determine AC discharge power. The PCS and grid connection constrain how quickly stored energy can reach the AC system, while the available SOC window constrains how long discharge can continue. More kWh can extend an energy-limited response without solving a power-limited one.

The research points to a connected explanation of dispatch value. Forecasts influence when energy is reserved or spent; feedback influences how the battery reacts to the present; equipment and reserve limits determine which reactions remain possible. The tariff then assigns a financial consequence to the resulting grid exchange. Improving one part can help, but its value depends on the others: a more accurate forecast cannot remove a power limit, and fast feedback cannot discharge energy that has already been used.

Model specifications and configuration information: PowerFac-261 technical documents · PowerFac-5M technical documents.