Self Organizing Neural Network for Adaptive Prediction of Dynamic Systems

本课题的目标是设计一种新颖的自组织神经系统，用于自适应预测受气候影响的每小时城市用水量干扰，并通过使用来自德克萨斯州阿灵顿市的实时序列来测试所提出的设计。它将应用多模型方法来进行神经识别。自适应预测系统由三个子网组成：(1)一层基于最小二乘的自适应预测因子（ape）；每一个都对应于先验系统输入输出模型集合中的一个模型，用于过滤测量和输入数据并执行多步超前预测；(2)中间层自组织自适应神经元在不进行统计假设检验的情况下进行模型切换；(3)自适应线性组合器（alc）的输出层，用于组合相同交货期的多个预测。自组织自适应预测器（SOAP）还采用各种在线预测评价反馈信号和其他非时间序列输入来自适应自组织自适应预测量权重以及提供强化信号来提高自组织子网的模型交换性能。​该项目的活动包括：1)开发SOAP系统，在计算机工作站上使用合成时间序列和模拟测试。离线训练、在线模式学习、收敛速度和连接权值的可塑性是仿真过程中的关键问题。2)使用来自德克萨斯州阿灵顿市的实时序列和操作数据，对经过微调的SOAP系统进行分析。本研究旨在推动神经识别的基础知识。通过在类人猿层实现ARMAX模型，消除了神经工程中对独立噪声的限制假设；关于现有的无法提前一步预测的另一个限制也通过使用基于系统的自适应预测算法得到缓解猿。

The goal of this project is to design a novel self- organizing neural system for adaptive prediction of hourly municipal water consumptions subject to climatic disturbances, and to test the proposed design by using real time series from the City of Arlington, Texas. It will apply a multiple-model approach to neuroidentification. The adaptive prediction system is composed of three subnetworks: (1) A layer of least-squares-based adaptive predictor elements (APEs), each of which corresponds to a model in the set of a priori system's input-output models, for filtering measurements and input data and performing multistep-ahead predictions; (2) A middle layer of self-organizing adaptive neurons that perform model switching without statistical hypotheses testing; (3) An output layer of adaptive linear combiners (ALCs) for combination of the multiple predictions of the same leadtimes. The self-organizing adaptive predictor (SOAP) also employs various on-line forecast evaluation feedback signals and other non- time-series inputs for adapting the ALC weights as well as providing reinforcement signals to enhance model switching performance of the self organization subnetwork. The activities of this project consist of: 1) Development of the SOAP system, using synthetic time series and simulation tests on a computer workstation. Off-line training, on-line pattern learning, convergence rates and plasticity of the connection weights are the key issues during the simulation. 2) Analysis of the fine-tuned SOAP system, using real time series and operation data from the City of Arlington, Texas. This research has is intend to advance the fundamental state of knowledge in neuridentification. The restricted assumption in neuroengineering on independent noises is removed by implementing the ARMAX model in the APEs layer; another limitation with respect to the existing inability to predict beyond one step ahead is also alleviated by the use of system- based adaptive prediction algorithms in the APEs.