基于循环神经网络的电能质量扰动识别

针对电能质量扰动信号识别算法复杂、识别率低等问题,提出一种将长短时记忆神经网络应用于电能质量扰动信号识别分类的新方法。首先在 Tensorflow中搭建长短时记忆神经网络,建立电能质量扰动信号分类模型。其次利用分类模型对电能质量扰动信号原始数据进行有监督学习,提取扰动信号的深层次特征,并将其连接到Softmax分类器输出各扰动信号的识别率。最后将电能质量扰动信号通过递归图生成的二维轨迹图像作为分类模型的输入,通过训练模型实现扰动信号的分类。仿真结果表明,该分类模型对电能质量扰动信号的一维和二维表示均有较好的分类准确率,可以有效识别7种单一扰动和6种复合扰动信号。     

Static VAR compensator using recurrent neural network

In this paper, an internal model control recurrent neural network method is used to control the switching of thyristor-controlled reactor in a static VAR compensator (SVC) system for regulating the voltage. The novel controller scheme contains several feedback loops instead of only a feed-forward loop as in the conventional recurrent neural network (RNN). In the proposed controller model, the RNN identifier creates a sample of the connected system and its output generates a part of inputs for the RNN controller which then sends the control signal to the SVC system. Three types of non-linear conditions are chosen to test the operational capability of the new control system to perform the voltage regulation satisfying the IEEE Std 519-1992. The test cases contain a three-phase fault power system, opening of one of the transmission lines in a double line transmission system and sudden changes in the load demand. Results show that the proposed control model is capable of regulating the voltage of the system in a desired range.     

Long-term load forecasting using improved recurrent neural network

In general, electric power companies must prepare power supply capability for maximum electric load demand because it is very difficult at present to store electric power. It takes several years and requires a great amount of money to construct power generation and transmission facilities. Therefore, it is necessary to forecast long-term load demand exactly in order to plan or operate power systems efficiently. Several methods have been investigated so far for the long-term load forecasting. However, because the electric loads consist of many complex factors, good forecasting has been very difficult. This paper proposes a long-term load forecasting method using a recurrent neural network (RNN). This is a mutually connected network that has the ability of learning patterns and past records. In general, when interpolation is used for unlearned data sets, the neural network provides reasonably good outputs. However, when extrapolation is used, such as in long-term load forecasting, some kind of tunings have been necessary to obtain good results. Therefore, to solve the problem, a method is proposed in which growth rates are used as input and output data. Using the proposed method, successful results have been obtained and comparisons have been made with the conventional methods.     

基于循环神经网络的PEMFC寿命预测

在长短期记忆(LSTM)循环神经网络(RNN)的基础上,通过减少控制门的数量,引入门控循环单元(GRU)RNN。利用质子交换膜燃料电池(PEMFC)在动态循环工况下的耐久性测试数据,训练、验证RNN模型,并对PEMFC的剩余使用寿命进行预测。基于GRU所得的预测结果,能准确跟随实际电压值的变化,在计算速度和准确度方面优于LSTM。在电流密度为0.71 A/cm2时,预测结果的均方误差可达0.0035。     

Distribution network state estimation based on attention-enhanced recurrent neural network pseudo-measurement modeling

Because there is insufcient measurement data when implementing state estimation in distribution networks, this paper proposes an attention-enhanced recurrent neural network (A-RNN)-based pseudo-measurement modeling metho. First, based on analyzing the power series at the source and load end in the time and frequency domains, a period-dependent extrapolation model is established to characterize the power series in those domains. The complex mapping functions in the model are automatically represented by A-RNNs to obtain an A-RNNs-based period-dependent pseudo-measurement generation model. The distributed dynamic state estimation model of the distribution network is established, and the pseudo-measurement data generated by the model in real time is used as the input of the state estimation model together with the measurement data. The experimental results show that the method proposed can explore in depth the complex sequence characteristics of the measurement data such that the accuracy of the pseudo-measurement data is further improved. The results also show that the state estimation accuracy of a distribution network is very poor when there is a lack of measurement data, but is greatly improved by adding the pseudo-measurement data generated by the model proposed.     

水下机器人的动态Petri递归神经网络控制方法

针对传统的水下机器人模糊神经网络控制器存在计算量大、抗环境扰动滞后等缺点,设计递归模糊神经网络控制器,通过在线的动态反馈增强水下机器人对环境变化的反应能力.并在网络的第三层即Petri层设计阈值,根据控制器误差的在线控制网络的学习和训练量,从而减少了模糊神经网络的计算量,提高反应速度.基于反向梯度传播原理,由能量函数设计了该网络的学习算法,并根据离散型李亚普诺夫函数确定了学习率参数,从而保证整个网络的收敛性.实验结果表明,该控制器能够提高递归神经网络的计算效率,减少控制误差,对外界干扰具有较强的鲁棒性,在水下机器人的控制方面取得了更好的效果.     

Dynamic behavioral modeling of nonlinear circuits using a novel recurrent neural network technique

In this paper, a new method called local-global feedback recurrent neural network (LGFRNN) is proposed for dynamic behavioral modeling of nonlinear circuits. The structure of the proposed method is based on recurrent neural network and constructed by time-delayed local and global feedbacks. Adding time-delayed feedbacks has a great impact on the learning capability of previous neural network-based methods. Moreover, time-delayed local feedbacks alleviate the problem of slow convergency of the conventional neural network-based methods in the training phase. The proposed LGFRNN can be trained only by having sampled input-output waveforms of the original circuit without knowing the internal details of the circuit. A training algorithm based on real-time recurrent learning (RTRL) is used to train LGFRNN. After the training phase, the proposed LGFRNN provides accurate macromodel of a nonlinear circuit. The proposed method is more accurate compared with the conventional neural-based models (which do not benefit from time-delayed local-global feedbacks) and also significantly reduces the training time of the conventional models. Moreover, proposed LGFRNN is faster than the existing models in simulation tools. The validity of the proposed method is verified by time-domain modeling of three nonlinear devices including commercial TI's SN74AHCT540 device, five-stage complementary metal-oxide-semiconductor (CMOS) receiver, and commercial TI's LM324 power amplifier.     

基于长短期记忆循环神经网络的AGC实时控制策略

大量新能源的接入以及电网中冲击负荷数量的剧增,使得电网对自动发电控制(AGC)策略提出了新的要求.简化AGC的一般控制流程,对比不同AGC策略的控制特性,在每个考核周期内选择控制效果更优的控制策略,并充分发挥多种控制策略在各自优势工况下的性能,以得到优秀控制数据集;在此基础上,以长短期记忆(LSTM)循环神经网络为神经...     

基于LSTM循环神经网络的短期电力负荷预测

为了保障电网安全稳定和电力市场高效运行,电网调度人员和电力市场参与者对电力负荷预测准确度提出了更高要求,分布式电源和间歇性负荷是影响负荷精准预测的关键因素。针对传统负荷预测方法无法同时对负荷本身变化规律及其影响因素进行建模的问题,提出基于长短期记忆单元(LSTM)的负荷预测方法。利用具备时序记忆功能的LSTM构建深度循环神经网络(RNN),综合考虑历史负荷和各类负荷影响因素建立负荷预测模型。该方法利用神经网络的特征提取能力和LSTM的时序记忆能力,能在更长的历史时间范围内辨识负荷内在变化规律及各类影响因素对负荷的非线性影响。基于实际负荷数据对不同历史时间窗口、不同网络架构的负荷预测性能进行验证,并与其他负荷预测算法进行比较,结果表明所提方法能有效提升负荷预测准确性。     

Robust speed-controlled induction motor drive based on recurrent neural network

This paper proposes a recurrent neural network speed controller for an induction motor drive. This speed controller consists of a recurrent neural network identifier (RNNI) and recurrent neural network controller (RNNC). The RNNI is used to provide real-time adaptive identification of the unknown motor dynamics. The RNNC is used to produce an adaptive control force so that the motor speed can accurately track the reference command. A back-propagation algorithm was used as the learning algorithm to automatically adjust the weights of the RNNI and RNNC in order to minimize the performance functions. The proposed control scheme can quickly estimate the plant parameters and produce a control force, such that the motor speed can accurately track the reference command. Both computer simulations and experimental results demonstrated that the proposed control scheme was able to obtain robust speed control.     

Inferring figure-ground using a recurrent integrate-and-fire neural circuit.

Several theories of early visual perception hypothesize neural circuits that are responsible for assigning ownership of an object's occluding contour to a region which represents the "figure." Previously, we have presented a Bayesian network model which integrates multiple cues and uses belief propagation to infer local figure-ground relationships along an object's occluding contour. In this paper, we use a linear integrate-and-fire model to demonstrate how such inference mechanisms could be carried out in a biologically realistic neural circuit. The circuit maps the membrane potentials of individual neurons to log probabilities and uses recurrent connections to represent transition probabilities. The network's "perception" of figure-ground is demonstrated for several examples, including perceptually ambiguous figures, and compared qualitatively and quantitatively with human psychophysics.     

融合边缘检测和递归神经网络的视频表情识别

为有效解决传统视频人脸表情识别通常只关注单张视频帧的空间特征,而忽略了相邻帧之间隐藏的时间特征的问题,提出一种结合边缘检测和递归神经网络的视频表情识别方法,利用梯度边缘检测准确地提取输入图像的纹理信息,同时提出一种分片交叉LSTM结构,提取出图像序列中隐藏的时空特征。实验在CK+和MMI视频库上进行,在OCNN-RNN网络中分别取得88.4%和69.7%的识别率,在GCNN-RNN网络中分别取得89.8%和73.6%的识别率,最终使用提出的加权随机搜索方法融合GCNN-RNN和OCNN-RNN两个网络之后,分别取得了94.6%和79.9%的识别率,均优于单流网络算法,证明了所提算法的有效性。     

动态对角递归网络在变压器故障诊断中的应用

介绍了动态对角递归网络,并针对BP算法收敛慢的缺点,提出了递推预报误差学习算法。利用该算法对神经网络的权值和域值进行训练,有效地提高神经网络的收敛性及增量学习能力。将动态对角递归网络应用到变压器的故障诊断中,利用改良三比值方法来实现诊断,建立了诊断的模型。利用部分数据进行了训练及故障诊断的仿真,结果表明了利用该方法进行变压器故障诊断的有效性。     

动态对角递归网络在变压器故障诊断中的应用

介绍了动态对角递归网络,并针对BP算法收敛慢的缺点,提出了递推预报误差学习算法.利用该算法对神经网络的权值和域值进行训练,有效地提高神经网络的收敛性及增量学习能力.将动态对角递归网络应用到变压器的故障诊断中,利用改良三比值方法来实现诊断,建立了诊断的模型.利用部分数据进行了训练及故障诊断的仿真,结果表明了利用该方法进行变压器故障诊断的有效性.     

Mirroring our thought processes [recurrent neural network and timeseries in forecasting]

To employ simple exponential smoothing in statistical forecasting, we essentially have to assume that the time series fluctuates at a gradually changing mean level. Forecasts are created on an iterative basis by weighing averages of observed values in the time series. The weights are assigned unequally with heavier weights applied to the most recent observations and exponentially declining weights to observations made far in the past. Yet, simple exponential smoothing still cannot help in making accurate predictions. One still has to monitor this forecasting system to determine whether or not the weights need to be adjusted to reduce forecasting errors. Since artificial neural network (ANN) technology provides us with weight adjusting algorithms, we propose using a special ANN architecture, a simple recurrent neural network. This network will provide a simple exponential smoothing forecasting system with an adaptive weighting scheme     

结合受限玻尔兹曼机的递归神经网络电力系统短期负荷预测

短期负荷预测的重要性随着电力企业的发展不断提高。传统的负荷预测虽然已经发展相对成熟,但现阶段对负荷预测的准确性要求逐渐提高。为满足发展需要,则要对现有的方法进行改进或建立新的预测方法。通过分析负荷预测数据周期性及周期内的特征,结合递归神经网络在分析时间序列数据的独特优势和受限玻尔兹曼机的强大的无监督学习能力,对结合受限玻尔兹曼机的递归神经网络的工作原理及训练过程进行了阐述。利用该网络进行了电力负荷数据预测实验验证并与其他神经网络进行了比较性实验。结果表明,所提出的神经网络较其他网络在电力短期负荷预测实验中有更高的准确性。     

供热过程时滞递推神经网络解耦器设计

针对供热过程质调一量调通道耦合特性和节能控制的需要,提出基于时滞递推神经网络的供热解耦方法.通过典型信号响应与最小二乘法结合的方法得到供热过程耦合系统模型,利用改进的假近邻法预估神经网络训练数据的嵌入维数,确定神经网络输入节点个数,引入时滞环节构建神经网络解耦器.采用时滞递推神经网络解耦器对供热耦合系统进行解耦,消除供热过程质调、量调通道间的非线性强耦合作用.仿真结果证明该方法具有良好的动态和静态解耦特性,能够满足供热过程多回路控制的要求,使供热系统能够跟踪节能设定值进行调节,实现供热节能和优质供热.     

基于深度卷积和门控循环神经网络的传感器运动识别

用于可穿戴传感器的人体运动识别任务的传统机器学习方法通常需要手工提取特征,可以自动提取人体运动数据特征的深度神经网络正成为新的研究热点。目前将卷积神经网络(CNN)和长短期记忆(LSTM)循环神经网络组合而成的DeepConvLSTM在识别精度方面有着优于其他识别方法的表现。针对带有长短期记忆循环单元的神经网络的训练较为困难的问题,提出了一种基于卷积神经网络和门控循环单元(GRU)的融和模型,并在3个公开数据集(ACT数据集、UCI数据集和OPPORTUNITY数据集)上与卷积神经网络和DeepConvLSTM进行了性能对比。实验结果显示,该模型在3个公开数据集上的识别精度都高于卷积神经网络,与DeepConvLSTM相当,但是收敛速度比DeepConvLSTM更快。