Fuel Cell State Estimation via Sliding Mode Observer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fuel cell observation systems face challenges in accurately measuring and controlling the purging of non-reactive species and water at the anode, particularly in real-time, due to restrictive hypotheses on pressure and temperature evolution, which affects the efficiency and reliability of fuel cell operation.
Innovation Solution
A method using a sliding mode observer that calculates an estimate of the electrochemical system's state by measuring electric current, voltage, anode and cathode pressures, and temperature, and applying a corrective term to minimize the difference between measurement and estimate vectors, allowing for real-time control of the purging process without requiring humidity sensors or restrictive assumptions on pressure and temperature variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an unscented Kalman filter is used to estimate the state of the fuel cell, then the state variables (quantity of nitrogen and water at the anode) can be estimated from measurable parameters, but restrictive hypotheses on the temporal evolution of pressure and temperature are required, which reduces the accuracy and applicability of the estimation
Solution Approach 1:
The patent replaces the unscented Kalman filter (a complex mathematical filtering system requiring restrictive hypotheses) with a neural network-based observer. The neural network learns the complex nonlinear relationships between measurable parameters (current, voltage, temperatures, pressures) and state variables without requiring explicit mathematical models or restrictive assumptions about pressure and temperature evolution. This substitution of the estimation mechanism eliminates the need for restrictive hypotheses while maintaining or improving estimation accuracy.
Solution Approach 2:
The patent changes the approach from model-based parameter estimation (requiring hypotheses on pressure and temperature temporal evolution) to data-driven parameter estimation using a neural network. The neural network is trained on operational data to learn the relationships between measurable parameters and state variables, allowing the system to adapt to varying operating conditions without requiring restrictive mathematical hypotheses about how pressure and temperature evolve over time.
2Productivity
If the purging of the anode distribution circuit is controlled based on state estimation, then the purging operation can be optimized, but the restrictive hypotheses on pressure and temperature evolution limit the reliability of the control system
Solution Approach 1:
The patent replaces the restrictive model-based control system with a neural network-based observer that does not require hypotheses on pressure and temperature evolution. The neural network directly estimates state variables from current measurements, providing more reliable control signals for the purging system across varying operating conditions. This substitution improves both the efficiency and reliability of the purging control by eliminating the fundamental limitation of the previous approach.
3Measurement precision
If direct measurement of nitrogen and water quantity at the anode is implemented, then real-time control of purging is possible, but such parameters are not directly measurable or require complex sensing systems
Solution Approach 1:
The patent creates a virtual copy of the difficult-to-measure parameters (nitrogen and water quantity) through neural network estimation. Instead of installing complex sensors to directly measure these state variables, the system uses a neural network to compute their values based on easily measurable parameters (current, voltage, temperatures, pressures). This copying approach provides real-time information about nitrogen and water quantity without the complexity of direct measurement systems.
Solution Approach 2:
The neural network acts as an intermediary that translates easily measurable parameters (current, voltage, temperatures, pressures) into estimates of difficult-to-measure parameters (nitrogen and water quantity). Rather than directly measuring the state variables, the system uses the neural network as a mediator to infer their values from the measurable parameters, avoiding the need for complex direct sensing systems.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
The invention relates to a method for observing a state (X) of an electrochemical system (1) comprising a fuel cell (2) comprising the following steps: a) measurement (100) of parameters representative of the fuel cell in operation; b) formation (200) of a control vector (U); c) formation (300) of a measurement vector (Y); d) calculation (400) of a so-called uncorrected time evolution ( ); e) calculation (500) of a correction term (ε) in sliding mode; f) calculation (700) of an estimate (X̂) of said state (X) of the electrochemical system (1); g) repetition of steps a) to f) by incrementing the measurement time.