Misproduction alarms notify you if there is a deviation between target and measured values, either on system or on inverter level. There are several methods for estimating the theoretical production (“simulation”) of the system under normal operating mode:
- Automatic inverter comparison
- Physical simulation model
- Machine learning-based simulation
If the deviation exceeds the threshold value that you have defined as “Normal”, an alarm is triggered. Whether the alarm severity is “High” or “Critical” depends on the thresholds you set up
The target value and the comparison to the measurement value are calculated on an interval basis both for the whole system and for each inverter each time the data is updated. You can define three different alarm severities based on the degree of deviation.
If the deviation changes over time, the severity of the alarm is automatically updated. A dummy tool in the VCOM user interface can help you to test the effect of different values on the alarm.
Simulation methods
The following methods are used to estimate theoretical production.
Method 1: Automatic inverter comparison
For this method, the target value source is determined by comparing the inverters with each other. The respective input configuration is automatically considered. For this, all inverters are grouped based on their input ratios. Then, the best performing inverter (the one with the highest normalized power) is established as a reference for other inverters in that group. In the next step, the normalized power is considered a target value for the inverters being compared.
Use case 1: Matching configurations
In this scenario, the VCOM Cloud would form two groups:
Group 1: Inverter 1 and Inverter 2 (1:1 ratio of 95°/20° and 275°/20°)
Group 2: Inverter 3 and Inverter 4 (1:1 ratio of 95°/30° and 275°/30°)
Use case 2: Multiple configurations that do not match another configuration
In this scenario, the VCOM Cloud would form three groups
Group 1: Inverter 1 and Inverter 2 (1:1 ratio of 95°/20° and 275°/20°)
Group 2: Inverter 3 and Inverter 4 (1:1 ratio of 95°/30° and 275°/30°)
Group 3: Inverter 5 and Inverter 6 (rest)
Use case 3: Single configuration that does not match any other configuration
In this scenario, the VCOM Cloud would form only one group containing all the inverters. Otherwise, there would not be any reference for Inverter 3.
Method 2: Physical simulation model
To gather the necessary data from the system configuration (the orientation and tilts of the system), the specific power (module and inverter) is first calculated for each orientation subsystem. The simulation is then generated based on the irradiance values for each orientation subsystem. Every subsystem is simulated with the irradiation from the sensor that is most similar to it. Similarity between subsystems and sensors is determined by calculating the correlation between the subsystem’s power measurements and the sensor radiation value.
- System configuration is retrieved, and the orientation subsystem is defined
- Installed sensors in the PV system are mapped to the respective orientation subsystem. Temperature values are taken into consideration
- Specific power for each orientation subsystem is calculated
- The values from the inverters and the installed power of the modules are used to calculate the specific yield for each orientation subsystem
- The target value per interval is calculated at the system level as a specific value
- The target value per interval is calculated at the system level as an absolute value
Method 3: Machine learning
Machine learning algorithms analyze the historical measurement data of the PV system and optimize the physical simulation. This reduces the deviation between the measured power and the target power to achieve the most accurate target value. The machine learning simulation method can help you to achieve the most accurate results. To apply machine learning, the system must meet the following criteria:
- 70% or more of the daytime data points are valid
- At least two weeks’ worth of valid training data within the last 30 days is available.
If valid training data is missing, the machine learning-optimized simulation will not be available, and a message is displayed on the simulation configuration page.