IEEE Power Electronics Magazine - March 2021 - 25

platform), component and converter data can be collected
and preprocessed. Real-time decision-making, such as fault
detection, protection, operation optimization, and stress
estimation, can be performed at this platform. Moreover,
models trained offline can be fine-tuned by online training [14] with converter-specific data. The edge platform is
locally implemented on the PES with resource-constrained
computation units. At the wind farm level (i.e., fog platform), data collected from individual converters can be
combined and analyzed for farm-level model training and
model updating. Applications such as RUL prediction and
degradation classification, which are relatively not timesensitive, can be performed at this platform. At the operator-level with multiple wind farms (i.e., cloud platform), a
massive amount of data are collected for model training.
The cloud platform is designed for computation- and dataintensive processing, which is generally equipped with
powerful computation units and data warehouse technology. Since it can be remotely accessed and controlled, there
is a less geographical requirement on this platform implementation. Maintenance plans requiring complex calculations can be formulated on this layer. It should be noted that
the fog platform is complementary to the cloud platform. It
is designed as an access point consisting of network gateways and data centers geographically closer to the edge
platform. It can coordinate the interoperability and allocate
resources between the units on the edge platform, enabling
a tightly connected structure. Meanwhile, the fog platform
is essential to system resilience in the case of interrupted
cloud connection.
For the hardware implementation, the cloud and fog
platform solutions can be extended from existing IIoT
applications in other fields with matured applications

Power
Consumption

Security

Flexibility

Latency

(e.g., Amazon Web Services, Microsoft Azure IoT). In
contrast, the existing implementation of AI tools on
the edge platform is limited by far. It requires embedded processors where AI tools can be implemented with
sufficient computation power and communication capability. Power consumption, security, flexibility, latency,
and cost are the most salient aspects of embedded processors. Figure 8 illustrates several off-the-shelf embeddable AI solutions for the edge platform with feature
comparisons. It covers several prevalent embedded solutions, including application-specific integrated circuit
(ASIC), field-programmable gate arrays (FPGA), microcontroller unit (MCU), digital signal processor (DSP),
graphics processing unit (GPU) and platform combination (e.g., MCU+FPGA). In [15], as an optimization tool, a
metaheuristic method real-coded jumping gene genetic
algorithm is applied to the parameters of the dynamic
model of the photovoltaic panel for data collection. The
tool is implemented with a hybrid platform of ARM MCU
and FPGA. In [16], a cloud-edge system is implemented
for the RUL prediction of MOSFETs in a power converter.
An ML tool LSTM is applied for the device degradation
modeling. It is implemented with NVIDIA Jetson TX2 for
the model inference on the edge layer. The RUL prediction can be completed within 26 ms, and the operational
power is less than 2 W.

Outlook on Data-Driven Condition
Monitoring for PES
The application of AI for data-driven condition monitoring
of PES is increasing in literature. While condition monitoring is demanded in critical power electronic applications,
the commercial implementation in power converters is

Price
ASIC

FPGA

Google TPU

XILINX Vitis AI Lattice sensAI Stack

MCU
NXP eIQ STM32Cube AI

Renesas e-AI

DSP
TI Sitara ML
GPU
NVIDIA Jetson TX2
Superior

Intermediate

Inferior

FIG 8 Comparisons of embeddable AI hardware platforms.

March 2021

z IEEE POWER ELECTRONICS MAGAZINE

IEEE Power Electronics Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Power Electronics Magazine - March 2021