Electricity price prediction method based on quantum complex neural network and Hilbert-Huang transform HHT

Abstract

The invention provides an original real-time electricity price prediction method based on a quantum complex neural network and Hilbert-Huang Transform (HHT), and the real-time electricity price prediction method based on the quantum complex neural network and the Hilbert-Huang Transform (HHT). Aiming at the non-stationarity of electricity price data on the historical level, firstly, the time internal correlation of each feature channel is extracted through an HHT time sequence analysis method, and complex multi-dimensional time sequence prediction is simplified into a simple regression task, so that a tedious time sequence modeling process is avoided; for the non-linear problem of electricity price data, a quantum neural network is used for capturing the coupling relation between different factors, and the calculation speed and the model efficiency are further improved by means of the parallelism of quantum calculation. Based on the two advantages, the method can realize accurate real-time electricity price prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the electricity price prediction problem in the field of new energy. First, the time internal correlation is extracted by calculating the instantaneous frequency through the classic Hilbert Huang Transform HHT, and then the correlation between different feature channels is modeled by combining quantum calculation, and the future electricity price is accurately predicted. BACKGROUND

[0002] Under the dual driving of global energy transformation and power market reform, electricity price prediction has become a key technology to support the stable operation of the power system and the efficient operation of the market. Nowadays, with a high proportion of renewable energy connected to the grid, electricity price stability is facing new challenges. Wind power and photovoltaic power generation are significantly affected by natural conditions, leading to frequent spikes and troughs in electricity prices. Superimposed with frequent extreme weather, the uncertainty of electricity prices is further aggravated. Accurate electricity price prediction is of great significance to all parties in the market: power generators can optimize their power generation plans, increase their output during high electricity prices to increase their profits; electricity retailers can use predictions to manage market risks and avoid losses caused by price fluctuations; and users can adjust their electricity consumption behavior based on the prediction to reduce electricity costs.

[0003] However, the current mainstream electricity price prediction techniques (such as LSTM, random forest) gradually show their limitations in dealing with complex situations. First, traditional models have difficulty effectively integrating unstructured information such as meteorological data, market strategies, and policy adjustments, making multi-factor coupling modeling increasingly difficult. Second, the lack of historical data results in insufficient model generalization ability, leading to low prediction accuracy in extreme scenarios. Finally, high-dimensional modeling requires a large amount of computing resources, making it difficult to meet the real-time needs of the spot market. SUMMARY

[0004] In order to more effectively solve the problems and challenges faced by the current electricity price prediction task, especially to capture multi-factor coupling relationships and improve the prediction efficiency of the model, the present application provides a real-time electricity price prediction method based on quantum complex neural networks and Hilbert Huang Transform HHT.

[0005] The technical scheme adopted by the present application to solve its technical problems is: firstly, the classical multi-dimensional time series electricity price data is preprocessed to ensure the stability of the numerical value and the convergence speed of the model. Secondly, the normalized electricity price data is divided into several time series signals according to the feature dimension, and the independent Hilbert-Huang transform (HHT) is performed to extract the instantaneous frequency. After HHT analysis of the internal correlation of time, a quantum complex neural network is constructed to model the complex nonlinear relationship between each feature dimension (excluding electricity price) (including classical layer coding, quantum gate circuit to realize quantum state evolution, and classical complex value layer decoding for final prediction output). Finally, the instantaneous frequency of each feature dimension after HHT is sequentially sent into the quantum complex neural network according to the time step to output the electricity price prediction value, and the mean square error function is combined to complete the regression analysis of the electricity price.

[0006] The beneficial effects of the present application are: first, the time internal correlation of each feature channel dimension is analyzed by HHT time series analysis, and the instantaneous frequency information obtained is directly used for regression analysis, which saves the cumbersome time series modeling process. Based on the analysis results of HHT, the nonlinear coupling relationship between different feature dimensions is mined by quantum computing, which improves the representation ability and calculation efficiency of the model, thereby realizing real-time prediction of complex multi-dimensional time series signals. BRIEF DESCRIPTION OF DRAWINGS

[0007] The present application will be further described below in conjunction with the drawings and examples.

[0008] Figure 1 is the electricity price prediction flowchart of the present application. DETAILED DESCRIPTION

[0009] I. Preprocessing operation For multi-dimensional time series electricity price data, the dimension is where T is the total time step, and C is the number of feature channels at each time step. In this task, we set C to 4, which includes temperature, power generation, power load and actual electricity price. For the specific electricity price data of each time step, we normalize each feature: ; where and are the normalized and unnormalized features, and are the minimum and maximum values of each feature in the overall time variation range. In this way, each feature is mapped to the [-1, 1] interval, ensuring the numerical stability of subsequent calculations. The first three are input features, and the electricity price feature is treated as a training label for the subsequent quantum complex neural network.

[0010] II. HHT time series analysis Because the electricity price data often presents irregular volatility in the historical level, the introduction of HHT has absolute advantages in the time series analysis of such non-stationary signals. We do independent HHT on the time series corresponding to each normalized feature dimension (temperature, power generation, power load) respectively, and the specific steps are as follows: (1) Empirical Mode Decomposition EMD: Since Hilbert Transform HT is only effective for narrowband signals, directly doing HT on non-stationary information will lead to the emergence of unexplained negative frequencies. Therefore, the goal of EMD is to decompose the original complex signal into a set of intrinsic mode functions IMF and a residual term, so that each IMF is an approximate "single component signal" to ensure that the calculated instantaneous frequency has physical meaning. The time series corresponding to the three input features (temperature, power generation, power load) ( ) are respectively decomposed by EMD to obtain: ; Among them and represent the kth IMF and the residual term of respectively. The energy of each IMF can be calculated by the square integral of its instantaneous amplitude . The dominant component is determined by the energy ratio , and is used for the next Hilbert transform HT: , ; (2) Calculate the instantaneous frequency: for each time series , the dominant component obtained by EMD decomposition is , which is simply denoted as , and its Hilbert transform is defined as: ; The dominant component and its Hilbert transform are combined into the analytic signal , and the instantaneous frequency can be obtained by taking the phase angle and differential operation of the analytic signal respectively: , ; The physical meaning of the instantaneous frequency is the instantaneous rate of change of the signal at a certain time, which avoids the limitation that the frequency is an interval average in Fourier transform, so it is more suitable for non-stationary signals such as electricity price data.

[0011] Three. Quantum complex neural network Since the instantaneous frequency information indirectly reflects the internal correlation of the time sequence signal (the stronger the time correlation, the more continuous and smooth the frequency change usually is; on the contrary, if the time correlation is weaker, the frequency change will present random fluctuations). Therefore, the next step is to further analyze the cross-dimensional correlation between the feature channels of the instantaneous frequency extracted by HHT, and we use quantum neural networks here, which are built as follows.

[0012] (1) Classical layer encoding: This part is a feedforward neural network structure that receives the result of the instantaneous frequency of the previous step HHT (i=1, 2, 3) as input (fed in time steps in turn) and encoded into a shallow feature vector: ; Where n is the number of neurons, which is set to 10 in this invention. The shallow feature vector O is not directly used for prediction, but as a regulatory signal for the subsequent quantum layer, so that the dynamic changes of the input data can be continuously captured by the quantum state.

[0013] (2) Quantum circuit: The quantum layer is the core component of the quantum neural network, which maps the parameters converted by the classical layer to the quantum state through quantum state initialization and quantum gate regulation, and naturally models the multi-dimensional correlation characteristics in the electricity price data. First, 10 qubits are used to ensure consistency with the number of neurons in the encoding layer, and are initialized to a pure state that is easy to regulate , the initial state of the system is represented by tensor product as: ; This initialization ensures that the initial state of each qubit is the same, and subsequent regulation only needs to map the differences in input data through rotation gates. If you need to expand, you can also set it to a uniform superposition state to further improve the information carrying capacity. Next, configure a rotation gate around the y-axis for each qubit , whose rotation angle is completely determined by the corresponding output of the previous classical layer, that is . For a qubit with an initial state of , the final state after rotation gate regulation is: ; The results show that the probability amplitude of the two ground states of the quantum state is directly determined by the input feature parameter , which shows that the data changes between feature channels are encoded into the probability distribution of the quantum state, realizing the transformation from classical to quantum.

[0014] (3) Classical layer decoding: The quantum state of the quantum layer cannot be directly used for subsequent prediction, and needs to be decoded into classical data through quantum measurement, and then processed by a classical complex neural network to finally obtain the predicted value of the electricity price. Specifically, we measure the probability of each qubit being in the state or The formula is: , ; Combine the two measurement results into a complex-valued observation vector: ; where n is the number of qubits, and the square root is to restore the probability amplitude information of the quantum state. Since the observation probabilities and are not independent, the ratio of the two directly reflects the size of the classical layer parameter , so the complex value encoding can more completely retain the probability correlation and phase information in the quantum measurement results. Finally, the complex vector is sent to the complex neural network for processing, and the modulus value of the output of the last layer is used for real-valued electricity price prediction.

[0015] Four. Training and optimization After constructing the quantum neural network, the instantaneous frequency information obtained by HHT time series analysis is sent into the quantum neural network in time steps. Each training data contains the frequency of different feature channels at this time node, and outputs the electricity price prediction value at the current time node. The mean square error function is used to compare the predicted value and the true value and minimize the regression error: .

Claims

1. A power price prediction method based on quantum complex neural network and Hilbert-Huang transform (HHT), which first calculates the instantaneous frequency characteristics of the time series corresponding to each power price influencing factor by HHT to extract the time correlation, and then combines quantum complex neural network to capture the coupling relationship between different factors to complete the regression prediction; characterized in that: The HHT is fused with the time sequence signal analysis capability, and the modeling and calculation advantages of the quantum complex neural network, and a complex multi-dimensional time sequence prediction problem is treated as a simple regression task. The steps include: (1) data preprocessing: the price data (including temperature, power generation, power load, and actual price) are normalized to map to the [-1, 1] interval to ensure numerical stability and model convergence speed; (2) HHT time sequence analysis: the features (including temperature, power generation, and power load) except the price are independently subjected to HHT on the time sequence axis, that is, the intrinsic mode function (IMF) with the maximum energy is screened out through empirical mode decomposition (EMD), and then the Hilbert transform is performed on the component to solve the instantaneous frequency information; (3) classical layer encoding: the instantaneous frequency information of each feature channel is sent to the classical neural network layer to encode into a shallow feature vector for subsequent quantum layer processing; (4) quantum circuit: the initialized quantum state is prepared and evolved through the quantum gate circuit, wherein the feature vector output by the classical layer is taken as the angle parameter of the quantum rotation gate, so that the change of the quantum state strictly follows the dynamic trend of the input data; (5) classical layer decoding: the probability of each quantum state being in the 0 ground state and the 1 ground state after evolution is measured to obtain a complex observation vector and send it to the complex neural network for processing to output the final price prediction value; (6) quantum neural network training: for the quantum complex neural network built in steps (3)-(5), a group of instantaneous frequencies at each time step is fed as a training data and the price prediction value is output, and then compared with the normalized actual price. The regression error is minimized by using the mean square error function, and the parameters are optimized to complete the training of the quantum complex neural network.

2. An original price prediction method based on quantum complex neural network and Hilbert-Huang transform (HHT), characterized in that the time sequence signal analysis capability of classical HHT and the powerful ability of quantum computing to process high-dimensional nonlinear data are fused; the HHT and quantum neural network are used to capture the correlation within the time and the correlation between the feature channels respectively, and the complex multi-dimensional time sequence prediction problem is simplified as a classical regression task; compared with the classical RNN and LSTM model, the method saves the cumbersome time sequence modeling process, and further improves the calculation speed and model efficiency by means of the parallelism of quantum computing, so as to realize accurate real-time price prediction.

Images