1        What are Reduxi AI strategies

Reduxi AI strategies are a set of intelligent, self-learning control algorithms embedded in the Reduxi EMS (Energy Management System). They forecast, optimise, and autonomously control energy flows across connected assets such as batteries, PV, EV chargers, and heat pumps.

The primary objective is profit optimisation.

1.1     Core definitions

• Forecast represents a prediction of future energy flows, including load, PV production, EV usage, and grid exchange. It is derived from historical measurements and external inputs such as weather conditions, time of day, and usage patterns.
• The objective is to optimise when to consume, store, or export energy. This includes, for example, charging the battery during periods of low prices or high PV production and discharging during high-price intervals. Electricity market prices (SPOT, ToU) are a key input to this process.
• The optimisation runs in rolling 15-minute intervals, continuously updating decisions based on the latest system state and forecasts.
• The AI strategy executes control indirectly by sending setpoints to the controller, which manages all included devices in real time.
• The system continuously learns from new data and updates its models and optimisation behaviour accordingly.

2        Initial configuration of AI strategies

2.1     Adding AI strategy

The AI strategy, or optimisation mode, is added in the Energy Manager strategy by clicking the “+” icon and selecting AI Energy Manager.

Configure the strategy parameters accordingly and save. More about the configurations can be found later in the device specific sections.

2.2     AI Strategy Conditions Checklist

To enable an AI strategy, the following conditions must be met:

• A strategy must be set with an AI optimisation mode, and only one AI strategy can be active at a time
• A price list must be configured
• Geolocation must be set
• Region, country, timezone, and currency must be defined
• The controller must be online and connected to the Reduxi cloud
• The devices must have relevant parameters configured, for example battery SOC, capacity, and power limits
• The location must have at least two weeks of historical data (PV, load, grid, etc.) so that reliable forecasts can be generated. See below for the behaviour without sufficient history.
• A valid AI licence must be active
• After all parameters are set, wait at least 15 minutes for the AI strategy to appear in the cloud app and start controlling devices
• A successfully configured AI strategy can be recognised by the forecast displayed to the right of the “Forecast” line. An example is shown below.

2.3     Advanced notes and options

• AI optimisation can run alongside default and schedule modes. This allows the system to behave differently at different times of day based on your settings. A typical example is to use different settings or limits during the day and at night.

• Negative price optimisation does not need to be configured separately. It is already handled within the AI optimisation mode when PV is configured accordingly.
• Grid limits need to be set so the system can reach the desired goal. For example, if the grid limit is set to 10 kW but the normal consumption is 100 kW, the optimisation target cannot be met and the user must define the limitation accordingly.
• If the AI optimisation cannot be achieved due to restrictive grid limits, the system automatically falls back to the configured default optimisation mode (e.g. self-consumption) as defined in the UI.
• Behaviour with limited historical data (First ~2 Weeks)
  • When there is not enough data for forecasting, no PV, load, or grid forecasts are available
  • In this case and if both import and export are allowed, the AI performs energy arbitrage
  • If either import or export is restricted, the AI switches to self-consumption mode
  • After enough historical data is available, the system automatically switches to regular AI mode
• Reduxi can control devices only to the extent that they correctly follow control signals. If a device does not respond as expected, its behaviour should be verified using manual control modes. For example, some PV power plants cannot be curtailed, some batteries do not follow the requested setpoint, and in hybrid inverter systems PV and battery control are often coupled. Reduxi adapts to these constraints, but the limitations still remain.
• Reduxi AI strategies coordinate all included devices towards a shared optimisation objective. Users can define which device categories are included, for example battery storage, and which are excluded, such as HVAC systems. We recommend all categories are included since this brings the higer value where everything works towards a common goal.
• Included devices are actively managed by the AI and participate in optimisation. Excluded devices operate independently under their own control logic and are not affected by the strategy.

2.4     Update rates:

• The AI strategy is recalculated every 15 minutes. At the same interval, the updated strategy is deployed to the controller, using the latest system state, weather data, and inputs.
• Once deployed to the controller, the strategy is executed locally in real time, reacting immediately to measurements and grid conditions.
• If a strategy is reconfigured, activation and initial forecast calculation take up to 15 minutes to become active.

3        Energy Consumption and Production Forecasting

Accurate forecasting of energy consumption and production is essential for effective AI strategy performance.

Using geolocation data, historical behaviour, and weather inputs (mainly solar irradiation and temperature), the Reduxi forecasting model predicts future energy flows for the next day and beyond.

The forecast is provided in 15-minute intervals and can be viewed by the user in the power chart, in 15-minute view.

3.1     Photovoltaic (Solar) Power Plant Production Forecast

Production of PV power depends on the historical behaviour, time of day and year, and weather input, mainly, irradiation, and temperature.

In order to simplify the configuration, Reduxi does not require the information of installed power, surface area, or orientation of the PV power plant. These characteristics are inferred automatically from historical PV production.

For an accurate forecast, a minimum of 14 days of historical data is required.

3.2     Consumption (Load) Forecast

Consumption forecast similarly uses historical data and weather inputs.

In addition, the model accounts for usage patterns such as time of day, for example lower consumption at night, and day of week, such as differences between weekdays and weekends.

Consumption forecasting is the most complex part, as it must account for less predictable behaviour within short timeframes, for example starting large machines in industrial environments or starting an EV charging.

At the moment, all loads are forecast together, including EV charging and HVAC. This will be further improved in future versions.

4        Battery Power Forecast

Battery power is the part of the system that can be actively controlled.

Its forecast is derived from optimisation goals such as peak shaving and cost optimisation. More details are provided in the following sections.

4.1     Grid Baseline Forecast

The grid-level forecast of consumption and production is essential for effective system control. It defines the expected energy exchange with the grid, and all AI optimisation is built around it.

The grid forecast is important because:

• Deviations from the expected baseline can lead to costs
• Consumption and production must stay within grid constraints, such as import and export limits or fuse limits

The baseline forecast enables the AI to:

• reduce grid import and use the battery instead
• shift consumption to manage peaks and avoid exceeding limits

For services such as flexibility or balancing, the baseline forecast also acts as the reference point because:

• it enables flexibility services as deviation from the baseline
• it determines potential penalties or rewards

5        Description of AI strategies per category

As mentioned, AI strategies can control all device categories to achieve a common objective. The following sections describe the functionality of each category in more detail.

5.1     AI strategy for battery (BESS)

Battery Energy Storage Systems (BESS) are used to maximise savings and earnings through the following modes:

• Energy arbitrage
  Charging the battery when energy prices are low and using or selling the stored energy when prices are higher.
• Charge from grid, discharge to load
  If discharge to the grid is not allowed, then the battery is discharged to cover the consumption. In this case, the battery is charged from the grid at the low-price intervals and used later when prices are higher. This shifts consumption and reduces costs.
• Charge from PV, discharge to grid
  If charging from the grid is not allowed, the battery is charged from PV and the energy can later be sold at higher prices. This increases the value of solar production compared to direct feed-in.

In all three modes, the Reduxi BESS AI module calculates the optimal time windows for charging and discharging. For this, it considers:

• Energy buying price, including all specified fees, network charges, and other components. (Note, the price must be defined in advance in the cloud app)
• Energy selling price, including all specified fees, network charges, and other components
• PV production, consumption, and grid forecast
• Grid and battery power limits
• Battery charge and discharge losses, as well as battery degradation cost

A few examples:

• Energy price is low at 12:00 and high at 20:00. The battery charges around 12:00 and discharges around 20:00.
• It considers grid and battery power limits. That means, if the grid limit is low and the required charging energy cannot be reached in a shorter time, the charging period is extended, for example from 11:00 to 15:00.
• It considers the PV production forecast and charges the battery from PV production instead of from the grid, or from a combination of both.
• If the price difference is too small to cover the charge and discharge losses and the battery degradation cost, the battery does not perform the charge/discharge cycle.

5.1.1    BESS AI configuration

The image shows a screenshot of the AI BESS configuration.

• Arbitrage or grid-restricted mode

The operating mode is determined by grid charge and discharge restrictions. Using the UI switches, the system can be configured for pure arbitrage or for optimisation that respects grid import and export constraints.

• Min savings/earnings per kWh

To account for battery degradation, the user can define the minimum savings or earnings that a cycle must generate. If the expected value is below this threshold, the battery is not used. This protects battery lifetime. A typical recommended value is 0.02 EUR/kWh, which has been estimated using the following calculation:

Cost of the battery / total energy over battery lifetime = 50,000 EUR / (10,000 cycles × 250 kWh) = 0.02 EUR/kWh

• Capacity reserved to cover load

As the forecast can never been ideal a percentage of the battery capacity can be reserved to cover local consumption and prevent exceeding grid limits.

• Capacity reserved for surplus (PV) production

Similarly, a percentage of the battery capacity can be reserved to PV surplus production in case the PV produces more than planned.

• The optimiser also considers energy loss during charging and discharging. The energy loss of the complete charge and discharge cycle is estimated at 8%.

5.2     AI strategy for PV

The PV AI strategy works best in combination with a battery. In this case, PV production can be stored in the battery and then reused in periods of high energy prices.

If the battery is not involved, still the PV AI strategy maximizes the profit. This means:

• when the energy price is positive, the PV production covers the load and any surplus can is exported to the grid up to the export limit
• when the sell price of energy is negative but the buy price is positive, the system works as self-consumption with zero export. In this state, the PV production covers the load but is not exporting to the grid
• when both the buy and sell prices of energy are negative, the mode is set to max-consumption. This means PV production is stopped, since it would otherwise result in losses

5.3     AI strategy for EV charging (EVSE)

TO BE DEFINED

5.4     AI strategy for heat pump control (HVAC)

The HVAC AI strategy is designed to reduce energy costs without compromising comfort. Buildings have high thermal inertia, which means heating and cooling can be shifted in time with only a small impact on indoor conditions.

When controlling heat pumps, Reduxi uses the device’s supported operating modes, such as OFF, ECO, NORMAL, and BOOST. This approach ensures the heat pump operates within its designed control logic, maintaining optimal efficiency and protecting the inverter. As a result, operation remains safe and does not negatively impact equipment lifetime.

The heat pump savings are achieved in two main ways:

a. Price-based optimisation
Energy prices vary throughout the day. Reduxi takes advantage of this by increasing heat pump operation when prices are low and reducing it when prices are high.

For example, if electricity is more expensive between 17:00 and 20:00, the system can pre-heat the building slightly before this period. During expensive hours, it reduces consumption while maintaining comfort through the building’s thermal inertia.

This simulated example shows how the heat pump increases energy use during low-price intervals and reduces consumption during high-price intervals. The dotted grey line represents the energy price.

b. Weather-based optimisation
Heat pumps operate more efficiently at higher outdoor temperatures. This is defined through the Coefficient of Performance (COP). Reduxi uses weather forecasts to shift heating to periods when conditions are more favourable. This improves efficiency and lowers costs, again without affecting comfort.

This simulated example shows how the heat pump reduces energy use during the night and increases power during warmer daytime hours to improve efficiency.

The dotted orange line represents the temperature forecast.

6        Examples of AI strategies in practice

Key information for the AI strategy is visualised in the Charts view, ideally using the power chart in 15-minute resolution. This view includes:

• Upper graph
  • historical power measurements of
    • yellow – PV production
    • green – site consumption
    • blue – battery power
    • orange line – grid exchange
  • forecast from the current time until to the end of the day
• In the bottom graph
  • red – energy purchase price, including all defined fees
  • green – energy selling price, including all defined fees
  • prices must be preconfigured in the cloud web application

6.1     Example: A typical example with a battery and PV

A – charging the battery during the night at low prices while keeping the grid limit at 240 kW

B – discharging the battery to the grid at high prices

C – using solar production to cover consumption at the location

D – charging the battery from forecast solar production during a low-price period (forecast)

E – discharging the battery to the grid at high prices (forecast)

6.2     Example: Utility scale PV (dedicated location to PV production)

A – discharging the battery to the grid at high prices

B – low-price period. Charging the battery purely from solar, with 0 kW at the grid, which means no import and no export

C – after the battery is full, exporting PV energy to the grid

D – discharging the battery to the grid during the three high-price intervals to maximise earnings

6.3     Example: Household location with zero feed-in

A – charging the battery from the grid while keeping the grid limit at 6 kW

B – discharging the battery to reduce grid import during high-price periods

C – charging the battery from the grid and solar, with a maximum grid exchange of 6 kW, in order to store energy for the expensive evening hours

D – discharging the battery during the expensive evening hours. No export to the grid is allowed due to the feed-in limit

6.4     Example 4: Utility PV

A – intervals in which PV energy is either exported or stored in the battery. This depends on the cost in 15-minute intervals

B – low-price period, purely charging the battery

C – after the battery is fully charged, exporting PV production to the grid

D – exporting the stored energy during high-price periods

6.5     Example: Consumption forecast of a production location with 7:00-15:00 working time

Typical load forecast of a company with working hours from 7:00 to 15:00.

Next to this, a typical solar production forecast is shown. These two forecasts are used to plan energy use accordingly.