Past research

Microgrid is a key enabling solution to future smart grids by integrating distributed renewable generators and storage systems to efficiently serve the local demand. However, due to the intermittent and uncertainty of distributed renewable energy, the reliability and economic operations of microgrid are facing increasing new challenges. Traditionally, economic dispatch issue is considered as solving an off-line or on-line optimization problem whose objective function is prior known. However, accurate and determined function expression is difficult to formulate, and wrong expression may result in waste of electricity cost and causing security issues. Thus, it is desirable to reformulate the economic dispatch problem, and solve it in a data-driven way. This paper proposes a data-driven energy management solution based on Bayesian-optimization-algorithm (BOA) for a single grid-connected home microgrid. The proposed solution formulates the optimization problem without a closedform objective function expression, and solves it using BOA based data-driven framework. The proposed solution is a kind of black box function sequential global optimization strategy, and does not require derivative operation on the objective function. Besides, it can also solve the microgrid operation and parameter prediction uncertainty. Simulation results demonstrate the effectiveness of the proposed solution.

The Lithium-ion batteries experience performance degradation during use through a complex interplay of physical and chemical processes, such as capacity fade and increased resistance, which may lead to energy storage system failures and even catastrophic loss. Battery state of health (SOH) is generally applied to assess the battery degradation level, while the remaining useful life (RUL) is employed to predict catastrophic events, extend life cycles, and schedule healthcare actions. However, the health assessment and RUL prognosis is particularly challenging in complex practical scenarios because: 1) degradation behaviors are quite difficult to model, as hybrid and coupled physical-chemical aging processing may occur in batteries, such as solid-electrolyte interface (SEI) layer growth, lithium corrosion, plating, and diffusion stress. 2) complex and coupled external causes may occur during the battery whole lifespan, such as inappropriate depth of discharge (DOD), current, temperature, and even mechanical stress, which can result in intricate degradation phenomena. 3) degradation phenomena such as capacity loss and resistance rise, is slow time-varying process, which thus require long-term observation, massive data storage, and heavy data processing. These effffects may also influenced by other factors, such as temperature and state-of-charge (SOC), making degradation modeling more challenging and complex.

We develop degradation models and parameter estimation/prediction methodologies to handle the aforementioned challenges: 1) We formulate a state space model to describe battery degradation dynamics using stachasitic processes. 2) We develop online/offline parameter identification and state estimation methods based on a combination of Maximum likelihood estimation and Bayesian filters enable accurate forecasting and computationally efficient modeling.

In order to ensure the battery work within its safety area and best working performance with an optimal management strategy, effective battery management system (BMS) must evaluate the current amount of energy stored in the battery, its power capability and health, which requires on-line estimation of state-of-charge/energy (SOC/E) and model parameters of the battery. However, managing a lithium-ion battery is a fundamentally challenging problem, because 1) the parameters of lithium-ion batteries may change along with the ambient temperature and degradation, which will result in time-varying model parameters. 2) However, the state constraints (0 ⩽ SOC ⩽ 1) and parameters constraints (impedance nonnegativity) are often neglected because they do not fit easily into the structure of the filters. 3) The trade-off between model complexity and accuracy of equivalent circuit model have not been well addressed in previous research. In order to improve the state estimation accuracy and robustness of Lithium-ion batteries under dynamic load and temperature conditions, we develop several different parameter estimation methodolagies based on the combination of equivalent circuit models and Bayesian filters. In this topic, we mainly develop simultaneous parameter identification and state estimation approaches to handle challenges.

• State and parameter simultaneous estimation based on recursive least squares and Unscented KF, improving adaptability
• State and parameter simultaneous estimation based on Extended KF and Particle filter, further improving accuracy
• Proposed constrained dual Bayesian filtering, improving convergence speed and accuracy
• Proposed dual filtering framework based on Subspace identification method and Particle filter, on-line determine battery model order

Understanding the dynamic behaviour of lithium-ion batteries is the first step for accurate monitor battery status. However, the lithium-ion is a typical nonlinear and environmentally sensetive dynamic system. Therefore, I began my research with hysteresis phenomena modeling (a special nonlinearity of OCV-SOC relationship) of LiFePO4 batteries. We develop a second order equivalent circuit model with one hysteresis state to modeling battery behaviors. The invariant imbedding method is proposed to identify battery parameters. Then, we develop a wavelet nerual network to map the nonlinear relationship between temperature, current, state of energy and voltage. A particle filter is proposed to estimate battery state of energy based on this wavelet nerual network model. At last, we develop a piecewise linearization model based on table look-up method to reduce the complexity caused by Jacobian matrix operation in Extended Kalman filter, unscented transform in Unscented Kalman filter, and resampling process in particle filter. On the basis of the linearized model, a linear Kalman filter can be easily implemented to estimate battery states.

• Nonlinear parameter estimation based on piecewise linearization model and Linear Kalman filter, reducing computation complexity
• Battery dynamic behavior modeling based on Wavelet neural network modeling and particle filter, improving robustness
• Battery hysteresis modeling based on Equivalent circuit model, improve model accuracy