2.1 Stochastic Uncertainty Analysis and Modelling on The Power Supply Side

Load side time domain model, and in the MATLAB through module modelling and mathematical modelling method, in turn build wind power generation, photovoltaic power generation, steel plant, electrified railway, electric vehicles, and select the appropriate node access IEEE30 node power system, establish the source load dual end of uncertain new energy power system model, as shown in Eq. (1), finally through the Newton-ravson method tide calculation each node state.

The new energy power system faces uncertainty challenges on both the source and load sides. On the power supply side, due to the inherent randomness and volatility of wind speed and solar radiation, the power supply capacity of a high proportion of new energy power systems is uncontrollable ^[20]. It is mainly to construct the new energy power system model needed in this paper, as Eq. (2), how to accurately describe the uncertainty of new energy power generation is crucial, the following will respectively to the wind, light distributed power output characteristics.

As shown in Eq. (3), using wind energy to drive the wind wheel blade rotation, and through the gear gearbox and the speed regulating mechanism, and then connected to the generator to make the generator run evenly can realize the conversion of wind energy to electric energy.

Wind speed is an important performance of wind energy, but due to the different geographical location of wind turbine installation and climatic conditions, wind speed has a strong randomness, as shown in Eq. (4), in order to describe its change characteristics. The Weibull distribution is often used to describe natural phenomena because of its flexible transformations and simple functions.

Two parameters of Weibull probability density function are as follows: photovoltaic output power is mainly affected by the light intensity, as shown in Eq. (5), has obvious spatial and temporal distribution characteristics, and every day, every year also has a big difference.

Solar panels convert light energy into electricity, as shown in Equation (6), and the panels themselves are affected by the temperature and the intensity of the light. The learning model can be expressed as follows:

2.2 Random Uncertainty in the Shock Load

As shown in Eq. (7), it will have a great impact on the power quality of the power system when it is connected to the power system. The main impact is to produce a large number of harmonics, cause voltage fluctuations and flicker, and produce negative order current.

As shown in Eq. (8), the mathematical model of electric arc furnace is established, and the model is established in MATLAB / Simulink. The arc energy balance relationship is as follows:

Electrified railway is the most important type of railway in contemporary times. Its electric energy is provided by the electric traction power supply system, which is mainly composed of traction substation and contact catenary. The electric locomotive is its load unit. The contact net provides 25kV alternating current to the electric locomotive, as shown in Eq. (9), and reduces the voltage to 1500V through the transformer to supply power to the motor.

The access of a high proportion of new energy makes the traditional power system gradually evolve into a new power system mainly based on new energy, which brings difficulties to the system state prediction. To this end, relevant scholars have carried out research on the state estimation of new energy power system, as shown in Eq. (10).

For including wind, light and other distributed power supply of new energy power system output uncertainty, li gang and others summarized the wind turbine fault diagnosis and state prediction of the difficult problem, puts forward a probability based on superposition denoising automatic encoder prediction method, solve the new energy output uncertain problems to the system state prediction.As shown in Eq. (11), a photovoltaic power generation combination prediction method based on chaos theory-set to realize accurate prediction of photovoltaic output.

The above research lays a foundation for the state prediction of new energy power system, and especially provides a reference method for the state prediction of power system with uncertain power supply. As shown in Eq. (12), with the access of impact loads such as new energy electric vehicles in recent years, the load randomness and volatility of the regional power grid are constantly increasing.

3.1 Adaptive-Volume Kalman Filter Algorithm

The regional operation of the new energy power system is affected by meteorology, the use of power electronic devices and the operation state of power generation equipment, which greatly restricts the application of traditional state estimation to the uncertain power system of source and load ^[21,22]. Xu and other scholars analysed the causes of anomalies of synchronous phasor data, summarized the feasibility of the data detection and repair sites used by the existing detection and repair institute, and summarized the research status of the detection and repair technology of synchronous phasor anomaly data from the local, regional and system levels ^[23,24]. Li and others proposed a data-driven robust disturbance identification method, which can effectively extract the robust temporal characteristics of the bad data in the power management unit. Liu and other scholars identified the input and output bad data through hypothesis testing, and corrected them to enhance the robustness of the power system ^[25,26]. IMelo and other foreign scholars have proposed a new correction method for medium and bad data in harmonic state estimation of power distribution system, and introduced the measurement calibration vector in the optimization model, which can identify and correct multiple bad data. R. Martinez-Parrales et al. comprehensively analyse the quality of the estimated states by using weighted least squares and maximum normalized residual test, to detect, identify, correct or eliminate coarse differences between multiple measurements under Gaussian and non-Gaussian noise ^[27,28].

Fig. 1. State estimation diagram of active distribution network.

Fig. 1 shows the state estimation diagram of the active distribution network. To obtain high position and speed accuracy, an extended Kalman filter was introduced to the GPS / DR combined navigation system. The EKF algorithm is simple, although simple and easy to operate ^[29]. It is widely used in industry, but it still has many disadvantages, such as the non-linear model, which is large and error-prone; ignoring Taylor expansion causes low estimation accuracy and poor model robustness; weak tracking ability to the mutation data; possible filter divergence. To address these problems, the 1999 S. Julier UKF, it is the nonlinear function probability density distribution, with a series of samples to approximate the posterior probability density, the main steps are: in the original state distribution according to a rule to select some sampling points, the mean and the covariance of the mean and the mean and the covariance, the sampling points into the nonlinear function, nonlinear function set of points, through the point set and observation equation can find the transformed mean and variance. The simulation results show that the adaptive strong tracking UKF can accurately diagnose sensor faults in the flight control system and quickly and smoothly detect state mutations. G. An improved trace Kalman filtering algorithm based on radial basis function neural networks was proposed by Zhang et al. This algorithm enables effective extraction of weak signal at low SNR.

3.2 State Estimation of Two-End Uncertain System Under Multiple Scenarios

Volume Kalman filter algorithm, although compared with the Kalman filter algorithm has good performance, but due to the error in the modelling process, Q and R set as constant may lead to reduce the prediction accuracy, so introduce the noise based on the memory index weighting algorithm, update calculation for every moment Q, in order to adapt to the variation of noise ^[30]. Fig. 2 is the algorithm diagram of the noise estimator. The BP neural network and other models can only process a single input separately, and the previous input and the latter input have no relationship. However, if you need to deal with the time series problems related to the previous input, these models are difficult to do, and the recurrent neural network can meet this need. R. Bai and other scholars use cognitive RF identification and RNN deep learning network to identify the book movement in the process of reading, improve the identification accuracy, so as to judge the degree of readers 'demand for books, and provide a basis for library book procurement and readers' personalized services.

Fig. 2. Noise evaluator algorithm diagram.

LSTM neural network algorithm has the following advantages over other algorithms: its cell state only requires linear summation operation, the gradient can easily move between the network without decay, the neural network can make the memory between the recent information and the long-time information, allowing the data to decide which information to keep and which to forget. Long and short-term memory neural networks with these advantages can better capture the information of the past and deal with the timing problems. Nanchang university Zhou Yi scholars considering the influence of noise data on the prediction results, combined with empirical mode decomposition, sinusoidal cosine algorithm and long short memory network design a photovoltaic power prediction model, discuss the input variables, algorithm parameters affect the prediction results, and through the comparison experiment, the method than other methods have higher prediction accuracy. A novel LSTM identification method was proposed by Shu et al. Residual LSTM with residual connections is introduced to learn features and improve inference on group activity classes.

4.1 Improved Combined Algorithm of ACKF-LSTM

Traditional state estimation algorithm mainly includes the least squares method, Wiener filtering, Kalman filter, etc., but in recent years, with the development of machine learning technology, based on the computer of neural network algorithm gradually become popular algorithm, neural network is composed of neurons. Learning of the input sequence is realized through the information transfer between multiple neurons to accurately get the relationship between input and output. The most basic component structure of the neural network is the artificial neurons. Multiple neurons are combined together through different rules to extract features from the input data to complete the learning process. Fig. 3 is a distributed parallel information processing graph. Until 2012, improved computing technology had made many complex operations cheap. With Alex Net as a symbol, the neural networks can develop rapidly. Neural network itself need not prior to determine the mapping relationship between the input and output of mathematical equations, only through their own training, learn some rules, given the input value is closest to the desired output value, therefore, theoretically, as long as there is enough historical data, through the neural network can predict any desired results.

Fig. 3. Distributed parallel information processing diagram.

The experimental results show that the improved BP neural network algorithm can successfully control the inverted pendulum, with a faster control response speed and good stability energy than other conventional controls. Models such as BP neural network can only handle a single input separately, and the former input and the latter input are completely unrelated. However, if you need to deal with the time series problems related to the previous input, these models are difficult to do, and the recurrent neural network can meet this need. Jiang and other scholars of Shanghai Electric Power University proposed a neural network wind speed prediction method based on seasonal index adjustment. Aiming at the nonlinear relationship between historical wind speed, it used the nonlinear fitting ability of neural network and adjusted the wind speed time series with seasonal index to predict, which can effectively improve the accuracy of wind speed prediction. R. Bai and other scholars use cognitive RF identification and RNN deep learning network to identify the book movement in the process of reading, improve the identification accuracy, so as to judge the degree of readers 'demand for books, and provide a basis for library book procurement and readers' personalized services.

4.2 Bad Power Grid Data Injection

The mentioned adaptive volume Kalman filter algorithm has good prediction accuracy with little voltage variation, can more accurately predict the change trend of voltage amplitude and phase under normal working conditions, However, this algorithm also has some limitations, it is difficult to guarantee the prediction accuracy of the adaptive volume Kalman filter algorithm for the mutation points with large voltage change. We propose an adaptive volume Kalman filter algorithm combined with long and short-term memory neural network algorithm. The basic structure of LSTM consists of input, output and forgetting gates, which can remember values at any time interval as well as information inside and outside the three gating units. Therefore, LSTM is suitable for time series analysis and time series-based data prediction. The subsequent research scenarios in this paper are to accurately predict the value of the mutation point based on the time series, including ultra-short time series prediction, such as LVRT, HVRT, and continuous LVRT to HVRT scenarios, as well as the relatively scattered predictions in the time series, mainly bad data injection scenarios. LSTM can effectively avoid the gradient vanishing and gradient explosion problems, with a strong generalization ability. Therefore, LSTM can solve such state estimation problem perfectly. Fig. 4 shows the adaptive volume Kalman filter algorithm diagram for the architecture of the proposed combined algorithm. The main steps are as follows: establish the regional new energy power system simulation model under different extreme conditions; using the conventional adaptive volume Kalman algorithm to predict the voltage amplitude in different scenarios; determine the position of mutation point or bad data in each scene; take the state amount before mutation data as feature input value, take the voltage value at the mutation data as output value; and select multiple sets of input and output values as sample set to train LSTM parameters and model. Get the accurate estimation of mutation data; substitute the exact estimate of mutation amount into the estimate of ACKF to obtain the state estimate of the whole system and compare it with the actual value.

Fig. 4. Adaptive volume kalman filter algorithm diagram.

A unique feature of photovoltaic power generation and wind turbines is that they usually use power electronic devices such as inverters to design interfaces rather than synchronous generators. At the same time, the level of power dispatching automation increases year by year, the measurement and acquisition device, transmission system and control system increases, and the overall electronic level of power grid increases. The large and complex power system causes the measuring equipment may not be the accurate data of the current system operation. Under normal circumstances, the probability of measuring data collected by the measurement system is 68.27%, the probability of $\pm 2 \sigma$ is 95.45%, and $\pm 3 \sigma$ is 99.73%. Therefore, the measurement data with error greater than $\pm 3\sigma$ is usually considered as bad data.