Extended Kalman Filter Learning Research Articles

Improving Microgrid Low-Voltage Ride-Through Capacity Using Neural Control

In this article, a neural sliding-mode linearization controller is proposed to regulate the generated active and reactive power for each distributed energy resource in a microgrid. The developed controller is based on recurrent high-order neural network identification, trained online with an extended Kalman filter learning algorithm. Based on such neural identification, adequate models of the microgrid generation units are obtained even in the presence of grid disturbances, which helps the proposed controller to reject disturbances, to ensure stability, and to operate the renewable energy sources under different grid scenarios. The proposed microgrid is composed of a wind power system, a solar power system, a battery bank, and a load demand. In addition, the microgrid under study is interconnected to an IEEE nine-bus system. The whole system is simulated in real time using the Opal-RT (OP5600) simulator. Real-time simulation results illustrate the effectiveness of the proposed control scheme to achieve trajectory tracking of the distributed energy resources active and reactive power even in the presence of grid disturbances.

Nonlinearities In Stock Return Prediction: Evidence From South Africa

<span style="font-family: Times New Roman; font-size: small;"> </span><p style="margin: 0in 35.7pt 0pt 0.5in; text-align: justify; mso-layout-grid-align: none; mso-outline-level: 1;" class="MsoNormal"><span style="font-family: Times New Roman;"><span style="color: black; font-size: 10pt; mso-themecolor: text1; mso-fareast-language: ZH-HK;">This research investigates the relationship between firm-specific style attributes and the cross-section of equity returns on the JSE Securities Exchange (JSE) over the period from 1 January 1997 to 31 December 2007. Both linear and nonlinear stock selection models are constructed based on the cross-section of equity returns with firm-specific attributes as model inputs.</span><span style="color: black; mso-themecolor: text1; mso-fareast-language: ZH-HK;"><span style="font-size: small;"> </span></span><span style="color: black; font-size: 10pt; mso-themecolor: text1; mso-fareast-language: ZH-HK;">Both linear and nonlinear models identify book-value-to-price and cash flow-to-price as significant styles attributes that distinguish near-term future share returns on the JSE.</span><span style="color: black; mso-themecolor: text1; mso-fareast-language: ZH-HK;"><span style="font-size: small;"> </span></span><span style="color: black; font-size: 10pt; mso-themecolor: text1; mso-fareast-language: ZH-HK;">The risk-adjusted performance of the nonlinear models is found to be comparable with that of linear models.</span><span style="color: black; mso-themecolor: text1; mso-fareast-language: ZH-HK;"><span style="font-size: small;"> </span></span><span style="color: black; font-size: 10pt; mso-themecolor: text1; mso-fareast-language: ZH-HK;">In terms of artificial neural network modeling, the extended Kalman filter learning rule is found to outperform the traditional backpropagation approach. This finding is consistent with our prior findings on global stock selection.</span></span></p><span style="font-family: Times New Roman; font-size: small;"> </span>

Application Of Cascade-Correlation Neural Networks In Developing Stock Selection Models For Global Equities

We investigate the potential of artificial neural networks (ANN) in the stock selection process of actively managed funds. Two ANN models are constructed to perform stock selection, using the Dow Jones (DJ) Sector Titans as the research database. The cascade-correlation algorithm of Fahlman and Lebiere (1990/1991) is combined with embedded learning rules, namely the backpropagation learning rule and the extended Kalman filter learning rule to forecast the cross-section of global equity returns. The main findings support the use of artificial neural networks for financial forecasting as an active portfolio management tool. In particular, fractile analysis and risk-adjusted return performance metrics provide evidence that the model trained via the extended Kalman filter rule had greater strength in identifying future top performers for global equities than the model trained via the backpropagation learning rule. There is no distinguishable difference between the performances of the bottom quartiles formed by both ANN models. The zero-investment portfolios formed by longing the top quartiles and simultaneously shorting the bottom quartiles or the market proxy exhibit statistically significant Jensens alpha and continues to accumulate positive returns over the out-of-sample period for both ANN models. On the other hand, the zero-investment portfolios formed by longing the bottom quartiles and simultaneously shorting the market proxy exhibit statistically significant Jensens alpha and continues to accumulate losses over the out-of-sample period for both ANN models. The implementation of the extended Kalman filter rule in training artificial neural networks for applications involving noisy financial data is recommended.

EKF-Based PI-/PD-Like Fuzzy-Neural-Network Controller for Brushless Drives

This paper presents the development of a fuzzy-neural-network (FNN) proportional-integral (PI)-/proportional-derivative (PD)-like controller with online learning for speed trajectory tracking of a brushless drive system. The design implements the novel use of the extended Kalman filter (EKF) to train FNN structures as part of the PI-/PD-like fuzzy design. The FNN structure has two parallel FNN PI-/PD-like controllers, each with four internal layers. EKF trains each FNN by modifying the weights and the membership function parameters. Thus, the proposed EKF-based architecture presents an alternative to control schemes employed so far. The objective is to replace the conventional PI-derivative (PID) controller with the proposed FNN PI-/PD-like controller with EKF learning mechanism. Comparisons of the algorithm performances provide evidence of improvement of the FNN PI-/PD-like controller over PID control. A test bench enables design implementation in the laboratory on hardware using a dSPACE DS1104 DSP and MATLAB/Simulink environment. Experimental testing results show that the proposed controller learns and robustly responds to a wide range of operating conditions in real time.

Neural Network-Based Market Clearing Price Prediction and Confidence Interval Estimation With an Improved Extended Kalman Filter Method

Market clearing prices (MCPs) play an important role in a deregulated power market, and good MCP prediction and confidence interval (CI) estimation will help utilities and independent power producers submit effective bids with low risks. MCP prediction, however, is difficult, since MCP is a nonstationary process. Effective prediction, in principle, can be achieved by neural networks using extended Kalman filter (EKF) as an integrated adaptive learning and CI estimation method. EKF learning, however, is computationally expensive because it involves high dimensional matrix manipulations. This work presents a modified U-D factorization method within the decoupled EKF (DEKF) framework. The computational speed and numerical stability of this resulting DEKF-UD method are significantly improved as compared to standard EKF. Testing results for a classroom problem and New England MCP predictions show that this new method provides smaller CIs than what provided by the BP-Bayesian method developed by the authors. Testing also shows that our new method has faster convergence, provides more accurate predictions as compared to BP-Bayesian, and our DEKF-UD MCP predictions are comparable in quality to ISO New England's predictions.

H/sub ∞/-learning of layered neural networks

Although the backpropagation (BP) scheme is widely used as a learning algorithm for multilayered neural networks, the learning speed of the BP algorithm to obtain acceptable errors is unsatisfactory in spite of some improvements such as introduction of a momentum factor and an adaptive learning rate in the weight adjustment. To solve this problem, a fast learning algorithm based on the extended Kalman filter (EKF) is presented and fortunately its computational complexity has been reduced by some simplifications. In general, however, the Kalman filtering algorithm is well known to be sensitive to the nature of noises which is generally assumed to be Gaussian. In addition, the H/sub /spl infin// theory suggests that the maximum energy gain of the Kalman algorithm from disturbances to the estimation error has no upper bound. Therefore, the EKF-based learning algorithms should be improved to enhance the robustness to variations in the initial values of link weights and thresholds as well as to the nature of noises. The paper proposes H/sub /spl infin//-learning as a novel learning rule and to derive new globally and locally optimized learning algorithms based on H/sub /spl infin//-learning. Their learning behavior is analyzed from various points of view using computer simulations. The derived algorithms are also compared, in performance and computational cost, with the conventional BP and EKF learning algorithms.