Spot electricity price prediction method and system based on multi-source data fusion

Abstract

The present application relates to the technical field of electricity price prediction, and in particular to a spot electricity price prediction method and system based on multi-source data fusion, which generates a fusion sequence by collecting and fusing multi-source data, and calculates an electricity price quality fluctuation value; performs wavelet transform based on the fluctuation value to obtain corresponding components; extracts and time-aligns each component to obtain corresponding original and interpolation sequences to determine a fidelity factor, and determines a component comparison coefficient in combination with a timeliness factor; adjusts wavelet reconstruction coefficients and an adjustment amplitude based on the component comparison coefficient and a residual fluctuation value to obtain a fusion feature sequence; inputs the fusion feature sequence into a model to generate an initial sequence; determines whether to correct the initial sequence and a correction amount based on the dispersion and oscillation intensity of the initial sequence to obtain a final sequence. The present application dynamically adjusts wavelet reconstruction coefficients and corrects an initial prediction sequence, avoids unreasonable oscillation of the prediction result, and effectively improves the accuracy, timeliness and anti-interference of the spot electricity price prediction result.

Description

Technical Field

[0001] This invention relates to the field of electricity price forecasting technology, and in particular to a method and system for forecasting spot electricity prices based on multi-source data fusion. Background Technology

[0002] Currently, spot electricity prices are easily affected by multiple factors such as load changes, weather conditions, and market trading behavior, exhibiting random and nonlinear characteristics that make forecasting difficult. Although methods for forecasting spot electricity prices exist, they still suffer from several problems: a single forecast output format makes it difficult to assess risk in trading decisions; in the face of drastic price fluctuations, forecast errors can be amplified, leading to forecast failure; and mixing data from multiple time scales makes it difficult for models to distinguish between random disturbances, cyclical patterns, and long-term trends, thus reducing forecast accuracy.

[0003] Chinese Patent Application Publication No. CN120952836A discloses a method for predicting spot electricity prices using multimodal data fusion. The related technical solution includes real-time acquisition of multimodal sequences such as wind and solar power output, load, and matching signals; based on instantaneous bandwidth adaptive partitioning, the multimodal sequences are divided into steady-state bundles and disturbance bundles and written into a component pool; confidence information is extracted from the error curve of each sequence in the component pool through short-time window reconstruction and written into a confidence mapping table; the confidence mapping table is corrected based on phase matching degree and amplitude abrupt change; high-confidence components are first superimposed according to the corrected confidence mapping table, and then low-confidence components are phase-scaling smoothed before superimposing to generate the first-round fusion curve; a reference curve is output using a load-driven fast reference model, and the difference between the reference curve and the first-round fusion curve is calculated to obtain a deviation vector; the confidence mapping table is refreshed using the deviation vector, and the superposition and smoothing steps are iteratively executed until the deviation vector converges to a predetermined limit; the final electricity price prediction curve is output, and the confidence evolution trajectory is written back.

[0004] However, the above technical solutions do not consider the time-sensitivity decay of data sources, which causes historical data to interfere with the prediction results; they also do not consider that the electricity price in this region is affected by the price fluctuations in neighboring regions, making it impossible to couple between regional markets; and thus, they adopt fixed fusion and prediction strategies, which makes the existing technical solutions lack the ability to adapt to different fluctuations. Summary of the Invention

[0005] To address this, the present invention provides a spot electricity price forecasting method and system based on multi-source data fusion, which overcomes the problems in the prior art where the forecast sequence oscillates violently due to the failure to consider the timeliness differences of multi-source data, the lack of inter-regional coupling, and the inability to adjust the proportion of each frequency band component based on data quality, resulting in low forecast accuracy.

[0006] To achieve the above objectives, this invention provides a spot electricity price forecasting method and system based on multi-source data fusion, comprising: Collect multi-source datasets, which include power grid load data, weather forecast data, renewable energy clearing data, regional tie line power flow data, and historical electricity price data; Data fusion is performed based on the multi-source dataset, and the electricity price quality fluctuation value is calculated based on the generated fused time series sequence. Based on the electricity price quality fluctuation value, wavelet transform is performed on the fused time series to obtain random fluctuation components, periodic fluctuation components, and long-term trend components. Each component is extracted and time-aligned to obtain the corresponding original sequence and interpolated sequence, and the corresponding frequency band fidelity coefficients are calculated. Based on the frequency band fidelity coefficient, the frequency band fidelity factor corresponding to each component is determined, and the component comparison coefficient is determined with the corresponding preset timeliness factor. Based on the corresponding component comparison coefficients, determine whether to adjust the wavelet reconstruction coefficients of each component, and determine the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation values ​​to obtain the fused feature sequence. The residual fluctuation values ​​are determined based on the variance of the difference sequence of the original sequence, and the adjustment coefficients include random fluctuation adjustment coefficients, periodic fluctuation adjustment coefficients, and long-term trend adjustment coefficients. The fused feature sequence is input into the electricity price prediction model to generate an initial prediction sequence; The local outlier degree and local oscillation intensity of the initial prediction sequence are extracted. Based on the local outlier degree, it is determined whether to modify the initial prediction sequence, and based on the local oscillation intensity, the modification amount of the initial prediction sequence is determined to generate the final prediction sequence.

[0007] Furthermore, the process of constructing a predictive electricity price model includes: Historical fused time series sequences are generated based on historical multi-source datasets, and historical electricity price quality fluctuation values ​​are calculated. Based on the historical electricity price quality fluctuation value being greater than or equal to a preset electricity price quality fluctuation value, a historical fusion feature sequence is generated. Based on the fact that the historical electricity price quality fluctuation value is less than the preset electricity price quality fluctuation value, a historical average sequence is determined to be generated. The historical fusion feature sequence or the historical average sequence is used as the input feature, and the historical electricity price at the corresponding time is used as the output label. The model is trained by a supervised learning algorithm to obtain the predicted electricity price model.

[0008] Furthermore, the process of determining whether to correct the initial prediction sequence based on local outlier degree includes: Collect local discrete points of the initial prediction sequence within the sliding window, and calculate the total number of local discrete points to obtain the local dispersion. Based on the fact that the local dispersion is greater than or equal to the preset local dispersion, the initial prediction sequence is determined to be corrected.

[0009] Furthermore, the process of determining the correction amount of the initial prediction sequence based on the local oscillation intensity includes: The variance is calculated based on several first-order differences in the initial prediction sequence within the sliding window to obtain the local oscillation intensity. Based on the fact that the local oscillation intensity is greater than or equal to the preset local oscillation intensity, it is determined to increase the smoothing filter intensity applied to the initial prediction sequence; Based on the fact that the local oscillation intensity is less than the preset local oscillation intensity, it is determined to reduce the smoothing filter intensity applied to the initial prediction sequence.

[0010] Furthermore, the increase in the smoothing filter intensity is determined based on the difference in oscillation intensity and the preset difference in oscillation intensity, and the increase in the smoothing filter intensity is positively correlated with the difference in oscillation intensity. The reduction range of the smoothing filter intensity is determined based on the oscillation intensity offset value and the preset oscillation intensity offset value, and the reduction range of the smoothing filter intensity is positively correlated with the oscillation intensity offset value.

[0011] Furthermore, the process of determining whether to adjust the wavelet reconstruction coefficients of each component includes: The corresponding frequency band fidelity coefficient is calculated based on the root mean square of the difference sequence corresponding to each component and the root mean square of the corresponding original sequence, wherein the difference sequence is determined based on the difference between several corresponding original sequences and interpolation sequences at the same time point. The component comparison coefficient is obtained by calculating based on the frequency band fidelity factor and the corresponding preset timeliness factor, wherein the frequency band fidelity factor is the frequency band fidelity coefficient; Based on the fact that the corresponding component comparison coefficient is less than the corresponding preset component comparison coefficient, the wavelet reconstruction coefficient of the corresponding component is determined and adjusted.

[0012] Furthermore, the process of determining the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation value includes: Based on the fact that the residual fluctuation value is greater than or equal to the preset residual fluctuation value, the corresponding adjustment coefficient is reduced. Based on the fact that the residual fluctuation value is less than the preset residual fluctuation value, it is determined to increase the corresponding adjustment coefficient.

[0013] Furthermore, the reduction magnitude of the adjustment coefficient is determined based on the fluctuation difference and the preset fluctuation difference, and the reduction magnitude of the adjustment coefficient is positively correlated with the fluctuation difference; The increase in the adjustment coefficient is determined based on the fluctuation offset value and the preset fluctuation offset value, and the increase in the adjustment coefficient is positively correlated with the fluctuation offset value.

[0014] Furthermore, the process of determining the wavelet transform of the fused time series based on the electricity price quality fluctuation value includes: The signal-to-noise ratio and periodicity of the fused time series are calculated, and the electricity price quality fluctuation value is obtained by calculation. Based on the fact that the electricity price quality fluctuation value is greater than or equal to the preset electricity price quality fluctuation value, it is determined that wavelet transform will be performed on the fused time series sequence.

[0015] This invention also provides a spot electricity price forecasting system based on multi-source data fusion, comprising: The data acquisition module is used to collect multi-source datasets, which include power grid load data, weather forecast data, renewable energy clearing data, regional tie line power flow data, and historical electricity price data. The quality assessment module is used to perform data fusion based on the multi-source dataset and to calculate the electricity price quality fluctuation value based on the generated fused time series. The time-frequency decomposition module is used to determine, based on the electricity price quality fluctuation value, to perform wavelet transform on the fused time series to obtain random fluctuation components, periodic fluctuation components and long-term trend components. The fidelity coefficient generation module is used to extract and time-align each component to obtain the corresponding original sequence and interpolated sequence, and to calculate the corresponding frequency band fidelity coefficient. The weighted fusion module is used to determine the frequency band fidelity factor corresponding to each component based on the frequency band fidelity coefficient, and to determine the component comparison coefficient with the corresponding preset timeliness factor. The coefficient adjustment module is used to determine whether to adjust the wavelet reconstruction coefficients of each component based on the corresponding component comparison coefficients, and to determine the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation value to obtain the fused feature sequence. The prediction module is used to input the fused feature sequence into the electricity price prediction model to generate an initial prediction sequence; The correction module is used to extract the local outlier degree and local oscillation intensity of the initial prediction sequence, determine whether to correct the initial prediction sequence based on the local outlier degree, and determine the correction amount of the initial prediction sequence based on the local oscillation intensity to generate the final prediction sequence.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Data fusion is performed using multi-source data to determine the wavelet transform to be applied to the fused time series sequence; the original and interpolated sequences corresponding to each component are obtained, and the corresponding frequency band fidelity coefficients are calculated; component comparison coefficients are determined based on the frequency band fidelity coefficients and a preset timeliness factor; the wavelet reconstruction coefficients of each component are adjusted based on the corresponding component comparison coefficients; the fused feature sequence is input into the electricity price prediction model to generate an initial prediction sequence; the local dispersion and local oscillation intensity of the initial prediction sequence are extracted to determine whether to correct the initial prediction sequence and the corresponding correction amount; thus, the electricity price prediction process can perceive data quality in real time, diagnose abnormal fluctuations, and recursively correct deviations, thereby significantly improving the accuracy, smoothness, and anti-interference ability of the prediction sequence, ultimately ensuring the stability and engineering practicality of short-term electricity price prediction in complex market environments.

[0017] Furthermore, this invention trains the electricity price prediction model based on historical multi-source datasets, ensuring that the input dimensions of the training phase and the real-time prediction phase are completely consistent. This eliminates the need to add additional input dimensions, reduces the number of parameters in time-series models such as Long Short-Term Memory Networks, accelerates training and inference speeds, and eliminates the need for manually setting thresholds for feature selection. This makes the model building process reproducible and significantly improves the robustness, training efficiency, and prediction accuracy of the electricity price prediction model.

[0018] Furthermore, this invention determines the local dispersion by counting the number of outliers within a statistical window, thereby locating unreasonable abrupt changes such as spikes and spikes in the initial prediction sequence and avoiding misjudgments caused by manually setting fixed thresholds or relying on global statistical indicators. When the dispersion is greater than or equal to the preset dispersion, a correction is triggered to ensure that adjustments are only made when necessary. The variance of several first-order differences within the window is calculated to determine the local oscillation intensity, thereby quantifying the severity of the sequence's jitter. When the oscillation intensity is too high or too low, the smoothing filter intensity is increased or decreased to effectively suppress unreasonable high-frequency oscillations or avoid excessive smoothing that leads to loss of details.

[0019] Furthermore, the present invention also determines the magnitude of the increase or decrease in the smoothing filter intensity based on the magnitude of the oscillation intensity difference or offset value, thereby matching the change in filter intensity with the degree of oscillation, avoiding over-adjustment or under-adjustment that may be caused by fixed step size adjustment, making the correction process converge smoothly, and the final output electricity price prediction result is more reliable.

[0020] Furthermore, this invention determines the component alignment coefficients based on the frequency band fidelity factor and the preset timeliness factor, and adjusts them when the component alignment coefficients are less than the preset component alignment coefficients, i.e. when the data quality is poor or the timeliness is low, to ensure that high-quality components remain unchanged and avoid information loss caused by blind adjustment; and calculates the residual fluctuation value to reflect the degree of local instability in the time alignment process, so that the adjustment range of the reconstruction coefficients matches the actual degree of distortion, avoiding insufficient or excessive correction due to the use of a fixed adjustment amount.

[0021] Furthermore, the present invention also determines the magnitude of the decrease or increase of the adjustment coefficient by positive correlation through the fluctuation difference or fluctuation offset value, and uses piecewise linear adjustment to smooth the change of wavelet reconstruction coefficient, preventing subsequent feature sequence oscillation caused by abrupt adjustment, and the final generated fused feature sequence has a higher signal-to-noise ratio and prediction stability.

[0022] Furthermore, this invention calculates electricity price quality fluctuation values ​​based on signal-to-noise ratio and periodicity, avoiding one-sided decisions based solely on signal-to-noise ratio or periodicity, thus making the decision results more reliable. Based on the comparison results of electricity price quality fluctuation values ​​being greater than or equal to preset electricity price quality fluctuation values, wavelet transform is determined to avoid forcibly decomposing meaningless spurious components due to excessive noise or unclear periodic characteristics, thereby ensuring the authenticity and reliability of the extracted random fluctuation, periodic fluctuation, and long-term trend components. Simultaneously, based on the comparison results of electricity price quality fluctuation values ​​being greater than or equal to preset electricity price quality fluctuation values, the original fused sequence is directly used for prediction, avoiding meaningless wavelet decomposition calculations without sacrificing prediction accuracy. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the spot electricity price forecasting method based on multi-source data fusion, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the logic of determining the adjustment amount of the random fluctuation adjustment coefficient based on the residual fluctuation value in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the logic of determining the initial prediction sequence correction based on local oscillation intensity in an embodiment of the present invention. Figure 4 This is a block diagram of a spot electricity price forecasting system based on multi-source data fusion, according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0025] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0026] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0027] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] Please see Figure 1 The diagram shown is a flowchart illustrating the spot electricity price forecasting method based on multi-source data fusion according to an embodiment of the present invention. The process in this embodiment includes at least the following steps: Step S1: Collect multi-source datasets, which include power grid load data, weather forecast data, renewable energy clearing data, regional tie line power flow data, and historical electricity price data. Step S2: Perform data fusion based on multi-source datasets, and calculate the electricity price quality fluctuation value based on the generated fused time series sequence; Step S3: Based on the electricity price quality fluctuation value, perform wavelet transform on the fused time series to obtain random fluctuation components, periodic fluctuation components, and long-term trend components. Step S4: Extract and time-align each component to obtain the corresponding original sequence and interpolated sequence, and calculate the corresponding frequency band fidelity coefficient. Step S5: Determine the frequency band fidelity factor corresponding to each component based on the frequency band fidelity coefficient, and determine the component comparison coefficient with the corresponding preset timeliness factor. Step S6: Determine whether to adjust the wavelet reconstruction coefficients of each component based on the corresponding component comparison coefficients, and determine the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation values ​​to obtain the fused feature sequence. The residual fluctuation values ​​are determined based on the variance of the difference sequence of the original sequence, and the adjustment coefficients include random fluctuation adjustment coefficients, periodic fluctuation adjustment coefficients, and long-term trend adjustment coefficients. Step S7: Input the fused feature sequence into the electricity price prediction model to generate the initial prediction sequence; Step S8: Extract the local outlier degree and local oscillation intensity of the initial prediction sequence, determine whether to correct the initial prediction sequence based on the local outlier degree, and determine the correction amount of the initial prediction sequence based on the local oscillation intensity, so as to generate the final prediction sequence.

[0029] Specifically, in step S1, grid load data, weather forecast data, renewable energy clearing data, regional interconnection power flow data, and historical electricity price data are collected from the grid dispatching system, weather forecast interface, renewable energy forecasting system, regional interconnection monitoring system, and power trading center, respectively. For example, the sampling interval for grid load data, renewable energy clearing data, regional interconnection power flow data, and historical electricity price data can be set to 15 minutes, and the sampling interval for weather forecast data can be set to 1 hour; linear interpolation is used to refine the sampling interval of the weather forecast data from 1 hour to 15 minutes.

[0030] The encrypted meteorological data and other data with a sampling interval of 15 minutes are unified into a 15-minute time grid to generate a data sequence. The box method is used to identify outliers and linear interpolation is used to correct them. For missing values ​​with no more than 3 consecutive missing points, linear interpolation is used, and for missing values ​​with more than 3 missing points, forward imputation is used.

[0031] The above processing yields a time-synchronized, complete, and anomaly-free multi-source dataset for subsequent steps.

[0032] In step S2, data fusion is performed based on multi-source datasets, and the electricity price quality fluctuation value is calculated based on the generated fused time series sequence. In this embodiment, grid load data, weather forecast data, renewable energy clearing data, regional tie-line power flow data, and historical electricity price data at the same time point are extracted from multi-source datasets and arranged according to a preset order to form row vectors. These row vectors are then stacked vertically in chronological order to obtain a fused time-series sequence. The preset order is: grid load data, weather forecast data, renewable energy clearing data, regional tie-line power flow data, and historical electricity price data.

[0033] For each column of the fused time series, i.e. each data source, the signal-to-noise ratio of the corresponding data source is calculated, and the average value is taken to obtain the signal-to-noise ratio.

[0034] Taking power grid load data as an example, the signal-to-noise ratio is determined based on the product of a conversion factor and a base-10 logarithm. The proper value of the logarithm is determined based on the ratio of the variance of the corresponding sequence to the variance of the difference sequence. The difference sequence is obtained by performing a first-order difference calculation on the sequence. The conversion factor is determined based on the physical quantity type of the signal and noise, and can be set to 10 for example.

[0035] Similarly, the signal-to-noise ratios (SNRs) for weather forecast data, renewable energy clearing data, regional interconnection line power flow data, and historical electricity price data are calculated separately, and the calculation methods are consistent with those for grid load data.

[0036] The overall signal-to-noise ratio is determined by averaging the signal-to-noise ratios corresponding to power grid load data, meteorological forecast data, renewable energy clearing data, regional interconnection line power flow data, and historical electricity price data.

[0037] Furthermore, taking grid load data as an example, the periodic significance rate is determined by dividing the peak energy by the sum of the peak energy and the average power of the background noise. The peak energy is determined by adding the power values ​​of the frequency corresponding to the maximum power and the power values ​​of the two adjacent frequency points. The maximum power is determined by ignoring the DC component in the power spectral density and based on the maximum value of the power spectrum in the remaining frequencies and its corresponding frequency. The average power of the background noise is determined by calculating the arithmetic mean of the power of all remaining frequency points after removing these three frequency points.

[0038] Similarly, the periodic significance rates corresponding to meteorological forecast data, renewable energy clearing data, regional interconnection line power flow data, and historical electricity price data are calculated separately, and the calculation method is consistent with the periodic significance rate of power grid load data.

[0039] The comprehensive periodic significance rate is calculated by averaging the periodic significance rates corresponding to power grid load data, meteorological forecast data, renewable energy clearing data, regional interconnection line power flow data, and historical electricity price data.

[0040] The electricity price quality fluctuation value is determined based on the product of the normalized signal-to-noise ratio factor and the comprehensive periodicity significance rate. It is used to characterize the signal-to-noise ratio and the prominence of the periodic component in the fused time series. That is, the higher the electricity price quality fluctuation value, the higher the signal-to-noise ratio and the more prominent the periodic component in the fused time series.

[0041] The normalized signal-to-noise ratio factor is based on 0 and the maximum value of the signal-to-noise ratio factor. The signal-to-noise ratio factor is based on 1 and the minimum value of the per-unit signal-to-noise ratio. The per-unit signal-to-noise ratio value is determined by the ratio of the difference between the overall signal-to-noise ratio and the historical lower limit to the difference between the historical upper limit and the historical lower limit. The historical lower limit and the historical upper limit are both obtained by statistical analysis of the overall signal-to-noise ratio during historical stable operating periods. For example, the historical lower limit can be set to 0dB and the historical upper limit to 30dB.

[0042] In step S3, wavelet transform is performed on the fused time series based on the electricity price quality fluctuation value to obtain random fluctuation components, periodic fluctuation components, and long-term trend components. In this embodiment, a preset electricity price quality fluctuation value is set in advance, and the comparison result with the electricity price quality fluctuation value determines the wavelet transform of the fused time series to obtain random fluctuation components, periodic fluctuation components and long-term trend components.

[0043] In one specific embodiment, if the electricity price quality fluctuation value is greater than or equal to the preset electricity price quality fluctuation value, it indicates that the signal-to-noise ratio of the fused time series is high and the periodic component is prominent. At this time, wavelet transform can effectively separate information from different frequency bands.

[0044] Therefore, wavelet transform is applied to the fused time series sequence to provide components with clear physical meaning for subsequent hierarchical prediction. This embodiment uses the db4 wavelet (Dobessi 4 wavelet) to perform 4-level discrete wavelet decomposition on each column of the fused time series sequence.

[0045] Taking power grid load data as an example, the random fluctuation component of power grid load is determined by performing an inverse wavelet transform while retaining only the first-level detail coefficients and setting the remaining coefficients to zero; the periodic fluctuation component of power grid load is determined by performing an inverse wavelet transform after adding the second and third-level detail coefficients point by point while setting the remaining coefficients to zero; the long-term trend component of power grid load is determined by performing an inverse wavelet transform while retaining only the fourth-level approximation coefficients. Among them, the first-level detail coefficients, second-level detail coefficients, third-level detail coefficients, fourth-level detail coefficients, and fourth-level approximation coefficients are all obtained by performing a four-level discrete wavelet decomposition of the power grid load data using the db4 wavelet.

[0046] Furthermore, the above decomposition and reconstruction were repeated for the four data sources: weather forecast, new energy clearing, regional interconnection line power flow, and historical electricity price, respectively, resulting in three corresponding components: random fluctuation component, periodic fluctuation component, and long-term trend component.

[0047] The total random fluctuation component is calculated by averaging the random fluctuation components from five data sources. Similarly, the average values ​​for the periodic fluctuation component and the long-term trend component are also calculated to obtain the total periodic fluctuation component and the total long-term trend component.

[0048] In one specific embodiment, if the electricity price quality fluctuation value is less than the preset electricity price quality fluctuation value, it indicates that the signal-to-noise ratio of the fused time series is low or the periodicity is weak.

[0049] Therefore, it is determined not to perform wavelet transform on the fused time series sequence. To unify the input dimension of the prediction model, the multi-source data values ​​at each time point in X are averaged to obtain an average sequence. At this point, the average sequence is directly used as the fused feature sequence, and the process jumps to step S7.

[0050] Specifically, a preset electricity price quality fluctuation value is determined based on the median of the electricity price quality fluctuation values ​​of historical data over the past 90 days, to characterize the benchmark quality of the current sequence under ideal conditions; wherein, the historical data is divided into multiple subsequences by natural days, and the electricity price quality fluctuation value of each subsequence is calculated in step S2, and the median of the calculation results of all subsequences is taken as the value, which can be set to 0.6 for example.

[0051] In step S4, each component is extracted and time-aligned to obtain the corresponding original sequence and interpolated sequence, and the corresponding frequency band fidelity coefficient is calculated. In this embodiment, the following operations are performed on the total random fluctuation component, the total periodic fluctuation component, and the total long-term trend component, respectively.

[0052] Taking the total random fluctuation component as an example, the component data is read and arranged in chronological order to obtain the random fluctuation sequence. Linear interpolation is then used to map the original random fluctuation sequence to a uniform 15-minute interval reference time axis to obtain the random fluctuation interpolated sequence.

[0053] The random frequency band fidelity coefficient is determined based on the maximum value between zero and the original fidelity coefficient. The original fidelity coefficient is calculated based on the difference between the baseline fidelity coefficient and the root mean square ratio. The baseline fidelity coefficient is determined based on the upper limit of ideal fidelity and can be set to 1 for example. The root mean square ratio is determined by dividing the root mean square of the random fluctuation difference by the root mean square value of the initial random fluctuation sequence and the sum of a minimum positive number. The minimum positive number, used to avoid a denominator of 0, can be set to 10 for example. -10 Among them, the root mean square of the random fluctuation difference is determined by taking the square root of the average of the squared differences between the interpolation results and the initial values ​​at all original time points; the root mean square of the initial random fluctuation sequence is determined by taking the square root of the average of the squared values ​​of all initial values.

[0054] Furthermore, the interpolation results at each original time point in the random fluctuation interpolation sequence are determined based on linear interpolation to avoid the situation where original time points containing the initial sequence are missing.

[0055] Furthermore, the above operation is repeated for the total periodic fluctuation component and the total long-term trend component to obtain the corresponding frequency band fidelity coefficients; wherein, the frequency band fidelity coefficients include random frequency band fidelity coefficients, periodic frequency band fidelity coefficients and long-term frequency band fidelity coefficients.

[0056] In step S5, the frequency band fidelity factor corresponding to each component is determined based on the frequency band fidelity coefficient, and the component comparison coefficient is determined with the corresponding preset timeliness factor. In this embodiment, the frequency band fidelity coefficients of the corresponding components are used as the frequency band fidelity factors of each component by an identity mapping method, so as to avoid introducing additional parameters without physical basis by using nonlinear transformation, thereby reducing the objectivity and reproducibility of the method.

[0057] The corresponding component comparison coefficients are determined based on the product of the frequency band fidelity coefficient, the preset timeliness factor, and the fluctuation coefficient corresponding to each component.

[0058] Furthermore, a timeliness factor is preset, which is determined based on the ratio of the average sampling interval of the corresponding original sequence to the minimum value of the average sampling interval. In this embodiment, since all data sources have been unified to a 15-minute grid and the three total components are obtained by arithmetic average, the sampling interval of the original sequence is 15 minutes; the minimum value of the average sampling interval can be set to 15 minutes for example.

[0059] The volatility coefficient is based on the physical contribution and stability differences of each component to electricity price forecasting. Among them, the long-term trend component reflects the macro trend of electricity prices, and the long-term volatility coefficient can be set to 0.6 for example; the periodic volatility component reflects deterministic patterns such as daily and weekly fluctuations, and the periodic volatility coefficient can be set to 0.3 for example; the random volatility component reflects noise and sudden factors, and has the worst stability, and the random volatility coefficient can be set to 0.1 for example.

[0060] In step S6, it is determined whether to adjust the wavelet reconstruction coefficients of each component based on the corresponding component comparison coefficients, and the adjustment coefficients of the wavelet reconstruction coefficients are determined based on the residual fluctuation values ​​to obtain the fused feature sequence; wherein, the adjustment coefficients include random fluctuation adjustment coefficients, periodic fluctuation adjustment coefficients and long-term trend adjustment coefficients.

[0061] In this embodiment, a corresponding preset component comparison coefficient is preset, and the component comparison coefficient is compared with the corresponding component comparison coefficient. Then, based on the comparison result, it is determined whether to adjust the wavelet reconstruction coefficient of each component.

[0062] Taking the random fluctuation comparison coefficient as an example, in one specific embodiment, if the random fluctuation comparison coefficient is greater than or equal to a preset random fluctuation comparison coefficient, it indicates that the data quality of the random fluctuation component is qualified, and its waveform fidelity and timeliness meet the requirements, and it will not pollute the fused feature sequence. Therefore, the wavelet reconstruction coefficients of the random fluctuation component are determined not to be adjusted, thereby keeping its original reconstruction coefficients unchanged, so that the information of the corresponding component can be completely preserved.

[0063] In one specific embodiment, if the random fluctuation comparison coefficient is less than a preset random fluctuation comparison coefficient, it indicates that the data quality of the random fluctuation component is poor, the waveform distortion is large, or the timeliness is insufficient, which may introduce noise. Therefore, the wavelet reconstruction coefficient of the random fluctuation component is determined, and the adjustment direction and magnitude are determined by the subsequent judgment result of the corresponding residual fluctuation value, thereby suppressing the interference of the random fluctuation component on the fused feature sequence.

[0064] Specifically, a random fluctuation comparison coefficient is preset, which is directly determined based on the engineering tolerance upper limit of random fluctuation components in the electricity price forecasting business. The comparison coefficient of random fluctuation components is usually set within 0.1 to ensure the dominance of long-term trends and cyclical fluctuations, while avoiding excessive impact of random fluctuations on forecast stability. Therefore, it can be set to 0.05 for example.

[0065] Taking the periodic fluctuation comparison coefficient as an example, in one specific embodiment, if the periodic fluctuation comparison coefficient is greater than or equal to a preset periodic fluctuation comparison coefficient, it indicates that the waveform fidelity of the periodic fluctuation component is high, and it can effectively reflect the peak-valley alternation pattern in the daily electricity price curve and the price difference between weekdays and weekends. Therefore, the wavelet reconstruction coefficient of the periodic fluctuation component is determined without adjustment, thereby keeping its original reconstruction coefficient unchanged, so that the periodic pattern information can be completely preserved.

[0066] In one specific embodiment, if the periodic fluctuation comparison coefficient is less than the preset periodic fluctuation comparison coefficient, it indicates that the waveform of the periodic fluctuation component is severely distorted, and its periodic characteristics are destroyed by noise or abnormal data. If used directly, it may introduce false periods into the prediction results. Therefore, the wavelet reconstruction coefficients for adjusting the periodic fluctuation component are determined, and the adjustment direction and magnitude are determined by the subsequent judgment results of the corresponding residual fluctuation values, thereby suppressing the interference of distorted periodic components on the fused feature sequence.

[0067] Specifically, a pre-set periodic fluctuation comparison coefficient is directly determined based on the engineering tolerance upper limit of the periodic fluctuation component in the electricity price forecasting business; among which, periodic fluctuation is one of the core driving factors of spot electricity price, and its comparison coefficient is usually set between 0.2 and 0.4, and can be set to 0.25 for example.

[0068] Taking the long-term trend comparison coefficient as an example, in one specific embodiment, if the long-term trend comparison coefficient is greater than or equal to the preset long-term trend comparison coefficient, it indicates that the waveform fidelity of the long-term trend component is high, which can truly reflect the macroeconomic trend of electricity prices on a seasonal scale or under policy cycles, providing a stable benchmark for prediction. Therefore, the wavelet reconstruction coefficient is determined not to adjust the long-term trend adjustment coefficient, thereby keeping its original reconstruction coefficient unchanged and ensuring that the trend information is completely preserved.

[0069] In one specific embodiment, if the long-term trend comparison coefficient is less than the preset long-term trend comparison coefficient, it indicates that the waveform distortion of the long-term trend component is large, and its macroscopic trend is distorted or truncated. Direct use may lead to a systematic deviation in the prediction results. Therefore, the wavelet reconstruction coefficients for adjusting the long-term trend component are determined. The adjustment direction and magnitude are determined by the subsequent judgment results of the corresponding residual fluctuation value, thereby suppressing the interference of the distorted trend on the fused feature sequence.

[0070] Specifically, a long-term trend comparison coefficient is preset, which is directly determined based on the engineering tolerance upper limit of the long-term trend component in the electricity price forecasting business; among them, the long-term trend component directly affects the baseline of the forecasting results, and its fidelity requirement is the highest. The comparison coefficient is usually set between 0.5 and 0.7, and can be set to 0.5 for example.

[0071] Please see Figure 2 As shown, it is a flowchart of the logic for determining the adjustment amount of the random fluctuation adjustment coefficient based on the residual fluctuation value in an embodiment of the present invention.

[0072] Furthermore, taking the random fluctuation component as an example, the adjustment coefficient of the corresponding wavelet reconstruction coefficient is determined based on the residual fluctuation value, which is determined based on the variance of the difference sequence of the original sequence of the corresponding random fluctuation component.

[0073] A corresponding preset residual fluctuation value is set in advance, and the value is compared with the residual fluctuation value to determine the adjustment range of the random fluctuation adjustment coefficient in the wavelet reconstruction coefficient.

[0074] In one specific embodiment, if the residual fluctuation value is greater than or equal to a preset residual fluctuation value, it indicates that the original sequence and the interpolated sequence of the random fluctuation component deviate significantly, resulting in severe time alignment distortion and potential noise introduction by the component. Therefore, it is determined to reduce the random fluctuation adjustment coefficient, thereby reducing the weight of the random fluctuation component in the fused feature sequence to suppress interference.

[0075] The reduced random fluctuation adjustment coefficient is determined based on the difference between the initial random fluctuation adjustment coefficient and the adjustment reduction, to effectively suppress the interference of distortion components. The initial random fluctuation adjustment coefficient is determined based on a baseline value under ideal unbiased conditions and can be set to 1 for example. The adjustment reduction is determined based on the product of the adjustment step size coefficient and the fluctuation difference. The fluctuation difference is determined based on the difference between the residual fluctuation value and a preset residual fluctuation value, to characterize the degree of deviation between the original sequence and the interpolated sequence and the degree of effective information carried by the random fluctuation component. The adjustment step size coefficient is determined based on the sensitivity requirement for residual fluctuation and can be set to 0.5 for example. If the calculated reduced random fluctuation adjustment coefficient is negative, it is set to 0.

[0076] In one specific embodiment, if the residual fluctuation value is less than a preset residual fluctuation value, it indicates that the original sequence and the interpolated sequence of the random fluctuation component match well, the time alignment fidelity is high, and the component information is reliable. Therefore, it is determined to increase the random fluctuation adjustment coefficient to enhance the weight of the random fluctuation component in the fused feature sequence in order to retain effective information.

[0077] The increased random fluctuation adjustment coefficient is determined based on the sum of the initial random fluctuation adjustment coefficient and the adjustment increment, to highlight the effective information of high-quality components. The adjustment increment is determined based on the product of the adjustment step size coefficient and the fluctuation offset value. The fluctuation offset value is determined based on the difference between the preset residual fluctuation value and the residual fluctuation value, to characterize the degree of fit between the original sequence and the interpolated sequence. If the calculated increased random fluctuation adjustment coefficient is greater than the preset adjustment coefficient upper limit, it is set to the preset adjustment coefficient upper limit. The preset adjustment coefficient upper limit is determined based on the need to prevent a single component from excessively dominating the fused feature sequence, and can be set to 2.0 for example.

[0078] Specifically, the periodic fluctuation component and the long-term trend component are judged and adjusted using the same method as the random fluctuation component to obtain the corresponding periodic fluctuation adjustment coefficient and long-term trend adjustment coefficient, respectively.

[0079] Specifically, the preset residual fluctuation value is directly determined based on the typical error range of the random fluctuation component in the power system time series data; wherein, provided that the input data has been normalized to the [0,1] interval by Min-Max, the residual fluctuation value is usually no more than 0.1 and is generally controlled within 0.07, and can be set to 0.06 for example.

[0080] In this embodiment, the wavelet reconstruction coefficients of each adjusted component and their corresponding components are used to determine the fusion feature sequence.

[0081] The fusion feature sequence is determined by multiplying the random fluctuation adjustment coefficient, the periodic fluctuation adjustment coefficient, and the long-term trend adjustment coefficient by their respective total random fluctuation component, total periodic fluctuation component, and total long-term trend component, and then summing them up.

[0082] Specifically, in step S7, the fused feature sequence is input into the predicted electricity price model to generate an initial predicted sequence; In this embodiment, when constructing the electricity price prediction model, a supervised learning algorithm is used for training based on a historical multi-source dataset. The trained model is the electricity price prediction model, which can preliminarily predict the initial electricity price sequence based on the real-time generated fused feature sequence. The supervised learning algorithm can, for example, be a Long Short-Term Memory network, Extreme Gradient Boosting, or Random Forest.

[0083] The specific steps involved in constructing a predictive electricity price model are as follows: (1) Training data preparation Historical data from the past 90 days was extracted from the multi-source dataset obtained in step S1. The multi-source dataset has been standardized to a 15-minute time grid, and missing value imputation, outlier correction, time alignment, and Min-Max normalization have been performed. Specifically: Input variables: For the above historical period data, the historical electricity price quality fluctuation value is calculated using the same method as the electricity price quality fluctuation value, and the historical fusion feature sequence or historical average sequence is determined based on the comparison between the historical electricity price quality fluctuation value and the preset electricity price quality fluctuation value. Output variable: The future electricity price at the corresponding moment, in yuan / MWh, for example, the electricity price 15 minutes from now.

[0084] (2) Feature engineering and model selection Model Selection: Given the complex nonlinear and long-term temporal dependencies between electricity prices and multi-source data, machine learning algorithms that can effectively capture temporal dependencies are preferred, including Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), and extreme gradient boosting. This embodiment preferentially uses a two-layer LSM network, with each layer containing 64 hidden units and a univariate time series with a single input dimension of 1.

[0085] (3) Model training and hyperparameter optimization The preprocessed dataset is divided chronologically, with the first 80% serving as the training set and the last 20% as the test set, in order to preserve the temporal sequence.

[0086] Training Process: The one-dimensional training feature sequence corresponding to the training set data is used as the input feature. Taking a Long Short-Term Memory (LSTM) network as an example, the input is a one-dimensional sequence window of 96 15-minute intervals over a past 24 hours, and the output is the predicted electricity price for the next 15-minute interval, using the historical electricity price at the corresponding time as the output label. The selected model is trained using the backpropagation algorithm, with the loss function being the minimization of the mean squared error or mean absolute error between the predicted and actual values. The Adam optimizer is used to iteratively update the weights and bias parameters in the network. The number of training epochs is set to 100, and an early stopping strategy is used to prevent overfitting.

[0087] Hyperparameter optimization: Grid search or Bayesian optimization strategies are used to optimize the key hyperparameters of the model. For example, for Long Short-Term Memory (LSTM) networks, the optimized parameters include 1–3 hidden layers, 2–128 neurons per layer, a learning rate of 0.001–0.01, a dropout rate of 0.1–0.5, and a batch size of 32–256. For extreme gradient boosting, the optimized parameters include 100–1000 trees, a maximum depth of 3–10, and a learning rate of 0.01–0.3. The optimal hyperparameter combination is selected through cross-validation or performance on the test set to ensure the model has good generalization ability and avoid overfitting or underfitting.

[0088] (4) Model validation and deployment After the model is trained, its performance is evaluated using a reserved test set. Evaluation metrics include: Root mean square error (RMSE) measures the square root of the sum of squares of the deviations between predicted and actual values, and is sensitive to larger errors. Mean absolute error (MAE) measures the average absolute deviation between a predicted value and the actual value. The coefficient of determination represents the degree to which the model explains the trend of electricity price changes. It ranges from 0 to 1, with a value closer to 1 indicating a better fit. Mean absolute percentage error measures the percentage of prediction error relative to the true value, facilitating comparisons between data of different magnitudes.

[0089] When the root mean square error of the model on the test set does not exceed a preset threshold and the coefficient of determination is greater than 0.85, the model is considered successfully trained and can be deployed as a predictive electricity price model. This model receives real-time generated historical fusion feature sequences or historical average sequences and outputs an initial prediction sequence. The 0.85 value is determined based on the ratio of the joint variance of the periodic and trend components in the electricity price sequence to the total variance.

[0090] In this embodiment, the fused feature sequence or average sequence is input into the electricity price prediction model, and an initial prediction sequence is generated using an iterative prediction method. In this embodiment, the electricity price prediction model uses a two-layer long short-term memory network. The input is a window of fused feature sequences from several past time steps, and the output is the electricity price prediction value for several time steps. Each forward calculation outputs the electricity price prediction value for the next time step.

[0091] The initial prediction sequence is based on the fusion feature sequence values ​​of the current time and 96 previous time points. It then sequentially predicts the electricity price forecast values ​​for the next 96 time points, corresponding to the next 24 hours to output a complete daily forecast.

[0092] In this embodiment, the initial prediction sequence consists of the prediction time and the initial predicted electricity price. The electricity price unit is yuan per megawatt-hour. For example, if the current time is 9:00 AM, after 4 iterations of prediction, the prediction sequence for the next hour is obtained, namely {(9:15, 350.2), (9:30, 355.0), (9:45, 348.5), (10:00, 360.1)}. That is, the electricity price at 9:15 AM is 350.2 yuan per megawatt-hour, the electricity price at 9:30 AM is 355.0 yuan per megawatt-hour, the electricity price at 9:45 AM is 348.5 yuan per megawatt-hour, and the electricity price at 10:00 AM is 360.1 yuan per megawatt-hour.

[0093] Specifically, in step S8, the local outlier degree and local oscillation intensity of the initial prediction sequence are extracted. Based on the local outlier degree, it is determined whether to modify the initial prediction sequence, and based on the local oscillation intensity, the modification amount of the initial prediction sequence is determined to generate the final prediction sequence.

[0094] In this embodiment, an initial prediction sequence is collected under a sliding window. In this embodiment, the sliding window can be set to 5, that is, each window contains 5 consecutive initial electricity price prediction values.

[0095] Local dispersion is determined based on the number of discrete points within the window; where discrete points are determined based on the prediction points in the initial prediction sequence where the absolute value of the difference between the predicted value and the mean is greater than three times the sample standard deviation.

[0096] A preset local dispersion is set in advance, and the comparison result with the local dispersion determines whether to correct the initial prediction sequence.

[0097] In one specific embodiment, if the local dispersion is less than a preset local dispersion, it indicates that there are no discrete points within the current sliding window, and the fluctuation of the initial prediction sequence within the local range is at a normal level. Therefore, it is determined not to modify the initial prediction sequence, thereby directly retaining the original prediction value and avoiding the loss of effective information due to over-smoothing.

[0098] In one specific embodiment, if the local dispersion is greater than or equal to a preset local dispersion, it indicates that there is at least one discrete point within the current sliding window, and the initial prediction sequence exhibits abnormal jumps or outliers locally. Therefore, it is determined to correct the initial prediction sequence to suppress the impact of abnormal spikes on the prediction results.

[0099] Specifically, the preset local dispersion is directly determined based on the upper limit of tolerance for local abnormal discrete points in the electricity price forecasting project; under normal forecasting conditions, the number of discrete points in each window does not exceed 1, and is generally controlled between 0 and 2, which can be set to 1 for example.

[0100] Please see Figure 3 As shown, it is a flowchart of the logic for determining the initial prediction sequence correction amount based on the local oscillation intensity in an embodiment of the present invention.

[0101] Furthermore, the local oscillation intensity is determined by variance calculation based on several first-order differences in the initial prediction sequence within the sliding window. A preset local oscillation intensity is set in advance, and the correction amount of the initial prediction sequence is determined by comparing the result with the local oscillation intensity.

[0102] In one specific embodiment, if the local oscillation intensity is greater than or equal to a preset local oscillation intensity, it indicates that the initial prediction sequence has severe local oscillations, large differences between adjacent points, and unreasonable abnormal jitter. Therefore, it is determined to increase the smoothing filter intensity applied to the initial prediction sequence to effectively suppress abnormal fluctuations and make the prediction sequence smoother and more stable.

[0103] The enhanced smoothing filter strength is determined based on the sum of the initial smoothing filter strength and the filter strength increment, in order to restore the stability of the sequence; wherein, the initial smoothing filter strength is determined based on the default starting strength when applying standard smoothing to the normal fluctuation prediction sequence, and can be set to 1 for example.

[0104] The filter intensity increment is determined based on the product of the intensity conversion factor and the oscillation intensity difference. The oscillation intensity difference is determined based on the difference between the local oscillation intensity and the preset local oscillation intensity to characterize the severity of the jitter in the predicted value within the window. The intensity conversion factor is directly determined based on the ratio between the time resolution of the predicted sequence and the typical amplitude of electricity price fluctuations. Taking a 15-minute resolution as an example, the normal fluctuation range of electricity price between adjacent moments usually does not exceed 5 yuan per megawatt-hour. In order to make the filter intensity increment 0.1 for every 1 yuan per megawatt-hour increase in the oscillation intensity difference, the intensity conversion factor can be set to 0.1 for example.

[0105] Furthermore, the intensity difference is determined based on the difference between the local oscillation intensity and the preset local oscillation intensity, and the local oscillation difference is fed back to step S6, that is, the preset component comparison coefficient corresponding to the next prediction process is increased based on the intensity difference to effectively suppress residual high-frequency oscillations.

[0106] Taking the random fluctuation comparison coefficient as an example, the preset random fluctuation comparison coefficient in the next prediction process is determined by the sum of the current preset random fluctuation comparison coefficient and the intensity difference multiplied by the increasing step size. The increasing step size is determined based on the response sensitivity to local oscillation intensity exceeding the limit, and can be set to 1.2 for example. The preset upper limit value of the random fluctuation comparison coefficient is determined based on the upper bound of the confidence interval of the random fluctuation component in historical stable operating periods, and can be set to 0.2 for example. If the calculated value exceeds the upper limit, the upper limit is used. The periodic fluctuation comparison coefficient and the long-term trend comparison coefficient can be adjusted similarly.

[0107] In one specific embodiment, if the local oscillation intensity is less than the preset local oscillation intensity, it indicates that the initial prediction sequence is locally stable with small differences between adjacent points. In this case, the sequence fluctuation pattern characterizes the true trend of electricity price changes. Therefore, it is determined to reduce the strength of the smoothing filter applied to the initial prediction sequence, thereby reducing unnecessary corrections, preserving the effective change details in the original prediction values, and avoiding information loss and prediction distortion caused by over-smoothing.

[0108] The reduced smoothing filter strength is determined based on the sum of the initial smoothing filter strength and the reduction in filter strength, in order to avoid information loss and prediction distortion caused by over-smoothing.

[0109] The filter intensity reduction is determined based on the product of the intensity conversion factor and the oscillation intensity offset value; wherein, the oscillation intensity offset value is determined based on the difference between the preset local oscillation intensity and the local oscillation intensity, so as to characterize the stationarity of the predicted sequence.

[0110] Specifically, the preset local oscillation intensity is directly determined based on the typical upper limit of the first-order difference variance of adjacent points within the sliding window. For example, with a 15-minute resolution, the normal difference in electricity prices between adjacent moments usually does not exceed 3 yuan per megawatt-hour, and the corresponding upper limit of the difference variance is about 9 (yuan / MWh)². Considering a certain margin, the preset local oscillation intensity can be set to 12 (yuan / MWh)² for example.

[0111] The final prediction sequence is composed of each prediction time and its corresponding final prediction value. The final prediction value is determined by the sum of the current contribution and the historical contribution. The current contribution is determined by the product of the smoothing coefficient and the current initial prediction value. The historical contribution is determined by the product of the historical weight and the final prediction value of the previous point. The historical weight is determined by the difference between the initial smoothing filter strength and the smoothing coefficient. The smoothing coefficient is determined by dividing the initial smoothing filter strength by the sum of the initial smoothing filter strength and the adjusted smoothing filter strength.

[0112] Please see Figure 4 As shown, it is a block diagram of the spot electricity price forecasting system based on multi-source data fusion according to an embodiment of the present invention.

[0113] This embodiment includes a data acquisition module, a quality assessment module, a time-frequency decomposition module, a fidelity coefficient generation module, a weight fusion module, a coefficient adjustment module, a prediction module, and a correction module.

[0114] The data acquisition module includes a power grid data interface unit, a meteorological data interface unit, a renewable energy data interface unit, a tie-line data interface unit, and an electricity price data interface unit. The power grid data interface unit collects power grid load data from the power grid dispatch automation system or smart meters, with a sampling interval of 15 minutes. The meteorological data interface unit obtains meteorological forecast data such as temperature, humidity, and wind speed from the meteorological forecast system or meteorological observation station, with a raw sampling interval of 1 hour. The renewable energy data interface unit obtains renewable energy clearing data such as wind power and photovoltaic power from the renewable energy power prediction system or the centralized control center. The tie-line data interface unit obtains regional tie-line power flow data from the phasor measurement unit or the regional power grid dispatch system. The electricity price data interface unit obtains spot electricity price data from the public interface of the power trading center or historical database. Each interface unit also includes a data cleaning subunit, used to remove outliers that exceed a reasonable range and mark missing data points; a data alignment subunit, used to encrypt meteorological data to 15 minutes using linear interpolation and unify all data to the same time grid; a missing value filling subunit, which fills in missing points using linear interpolation or forward filling; and an outlier correction subunit, which uses box-line method to identify outliers and corrects them using linear interpolation.

[0115] The quality assessment module is connected to the data acquisition module and includes a completeness calculation unit, a signal-to-noise ratio (SNR) calculation unit, a periodicity significance calculation unit, and a quality index fusion unit. The completeness calculation unit calculates the proportion of valid data points in the fused time series. The SNR calculation unit calculates the ratio of the variance of each data source to the variance of the difference series as the argument of the logarithm, and determines the value based on the product of the conversion factor and the base-10 logarithm. The periodicity significance calculation unit performs a fast Fourier transform on each data source, extracts the ratio of the peak energy to the average power of the background noise, and takes the average value to obtain the periodicity significance. The quality index fusion unit determines the electricity price quality fluctuation value based on the comprehensive SNR and the comprehensive periodicity significance, compares it with the preset electricity price quality fluctuation value, and outputs a decision signal on whether to perform wavelet transform.

[0116] The time-frequency decomposition module is connected to the quality assessment module and includes a wavelet decomposition unit, a component reconstruction unit, and a component fusion unit. After receiving the decision signal for wavelet transform, the wavelet decomposition unit performs a 4-level discrete wavelet decomposition on each data source of the fused time series using db4 wavelet to obtain the detail coefficients of each level and the highest level approximation coefficients. The component reconstruction unit inversely transforms the first level detail coefficients into random fluctuation components, adds the second and third level detail coefficients and inversely transforms them into periodic fluctuation components, and inversely transforms the highest level approximation coefficients into long-term trend components. The component fusion unit performs an arithmetic average of the same type of components from the five data sources to obtain the total random fluctuation component, the total periodic fluctuation component, and the total long-term trend component.

[0117] The fidelity coefficient generation module is connected to the time-frequency decomposition module. It includes an initial sequence extraction unit, an interpolation sequence generation unit, and a fidelity coefficient calculation unit. The initial sequence extraction unit reads the original sequence of each component from the output of the inverse wavelet transform. The interpolation sequence generation unit uses linear interpolation to map the original sequence to a 15-minute equally spaced reference time axis to obtain the interpolated sequence. The fidelity coefficient calculation unit compares the original value with the interpolation result at the original time point to obtain the corresponding frequency band fidelity coefficient.

[0118] The weighted fusion module is connected to the fidelity coefficient generation module, and includes a frequency band fidelity factor determination unit, a timeliness factor calculation unit, and a comparison coefficient calculation unit. The frequency band fidelity factor determination unit uses an identity mapping to use the frequency band fidelity coefficient as the frequency band fidelity factor. The timeliness factor calculation unit calculates the average sampling interval of the original sequence of each component and normalizes it to obtain the timeliness coefficient. The comparison coefficient calculation unit calculates the comparison coefficient of each component based on the frequency band fidelity factor and the timeliness coefficient.

[0119] The coefficient adjustment module is connected to the weight fusion module and the correction module, and includes an adjustment judgment unit, a residual fluctuation calculation unit, and a reconstruction coefficient update unit. The adjustment judgment unit compares the comparison coefficients of each component with the preset component comparison coefficients to determine whether adjustment is needed. The residual fluctuation calculation unit determines the residual fluctuation value based on the variance of the difference sequence of the original sequence. The reconstruction coefficient update unit determines the adjustment range of the adjustment coefficients based on the difference or offset value between the residual fluctuation value and the preset residual fluctuation value, updates the wavelet reconstruction coefficients, and then multiplies the updated coefficients with the corresponding total components and sums them to obtain the fused feature sequence.

[0120] The prediction module is connected to the coefficient adjustment module and includes a model input unit and an initial prediction generation unit. The model input unit organizes the fused feature sequence into an input tensor using a sliding window. The initial prediction generation unit performs forward calculations using a pre-trained electricity price prediction model and outputs the electricity price prediction values ​​for multiple future time points to form an initial prediction sequence.

[0121] The correction module is connected to the prediction module and the coefficient adjustment module, and includes a local outlier calculation unit, a local oscillation intensity calculation unit, a smoothing intensity adjustment unit, a sequence correction unit, and a feedback parameter adjustment unit. The local outlier calculation unit traverses the initial prediction sequence with a sliding window, calculates the mean and standard deviation within the window, and counts the number of discrete points to obtain the local outlier. The local oscillation intensity calculation unit calculates the variance of the first difference between adjacent points within the window to obtain the local oscillation intensity. The smoothing intensity adjustment unit compares the local outlier with the preset local outlier to determine whether correction is needed, and further determines the adjustment range of the smoothing filter intensity based on the comparison result between the local oscillation intensity and the preset local oscillation intensity. The sequence correction unit uses the exponential smoothing formula to correct the points that need correction, generates the final prediction sequence, and outputs it to a visualization display device or a power trading decision system. The feedback parameter adjustment unit calculates the intensity difference based on the local oscillation intensity and feeds the intensity difference back to the coefficient adjustment module to adjust the preset component comparison coefficient in the next prediction process.

[0122] All technologies not mentioned in the above embodiments are existing technologies. It is understood that no specific limitation is made to any preset parameter or critical parameter in the embodiments of the present invention, and the above values ​​are not limited thereto. Those skilled in the art can adjust the preset parameters or critical parameters accordingly based on actual needs, analysis of historical data, or equipment usage.

[0123] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A spot electricity price forecasting method based on multi-source data fusion, characterized in that, include: Collect multi-source datasets, which include power grid load data, weather forecast data, renewable energy clearing data, regional tie line power flow data, and historical electricity price data; Data fusion is performed based on the multi-source dataset, and the electricity price quality fluctuation value is calculated based on the generated fused time series sequence. Based on the electricity price quality fluctuation value, wavelet transform is performed on the fused time series to obtain random fluctuation components, periodic fluctuation components, and long-term trend components. Each component is extracted and time-aligned to obtain the corresponding original sequence and interpolated sequence, and the corresponding frequency band fidelity coefficient is calculated. Based on the frequency band fidelity coefficient, the frequency band fidelity factor corresponding to each component is determined, and then compared with the corresponding preset timeliness factor. Based on the corresponding component comparison coefficients, determine whether to adjust the wavelet reconstruction coefficients of each component, and determine the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation values ​​to obtain the fused feature sequence. The residual fluctuation values ​​are determined based on the variance of the difference sequence of the original sequence, and the adjustment coefficients include random fluctuation adjustment coefficients, periodic fluctuation adjustment coefficients, and long-term trend adjustment coefficients. The fused feature sequence is input into the electricity price prediction model to generate an initial prediction sequence; The local outlier degree and local oscillation intensity of the initial prediction sequence are extracted. Based on the local outlier degree, it is determined whether to modify the initial prediction sequence, and based on the local oscillation intensity, the modification amount of the initial prediction sequence is determined to generate the final prediction sequence.

2. The spot electricity price forecasting method based on multi-source data fusion according to claim 1, characterized in that, The process of building a model for predicting electricity prices includes: Historical fused time series sequences are generated based on historical multi-source datasets, and historical electricity price quality fluctuation values ​​are calculated. Based on the historical electricity price quality fluctuation value being greater than or equal to a preset electricity price quality fluctuation value, a historical fusion feature sequence is generated. Based on the fact that the historical electricity price quality fluctuation value is less than the preset electricity price quality fluctuation value, a historical average sequence is determined to be generated. The historical fusion feature sequence or the historical average sequence is used as the input feature, and the historical electricity price at the corresponding time is used as the output label. The model is trained by a supervised learning algorithm to obtain the predicted electricity price model.

3. The spot electricity price forecasting method based on multi-source data fusion according to claim 1, characterized in that, The process of determining whether to correct the initial prediction sequence based on local outlier degree includes: Collect local discrete points of the initial prediction sequence within the sliding window, and calculate the total number of local discrete points to obtain the local dispersion. Based on the fact that the local dispersion is greater than or equal to the preset local dispersion, the initial prediction sequence is determined to be corrected.

4. The spot electricity price forecasting method based on multi-source data fusion according to claim 3, characterized in that, The process of determining the correction amount of the initial prediction sequence based on the local oscillation intensity includes: The variance is calculated based on several first-order differences in the initial prediction sequence within the sliding window to obtain the local oscillation intensity. Based on the fact that the local oscillation intensity is greater than or equal to the preset local oscillation intensity, it is determined to increase the smoothing filter intensity applied to the initial prediction sequence; Based on the fact that the local oscillation intensity is less than the preset local oscillation intensity, it is determined to reduce the smoothing filter intensity applied to the initial prediction sequence.

5. The spot electricity price forecasting method based on multi-source data fusion according to claim 4, characterized in that, The increase in the smoothing filter intensity is determined based on the difference in oscillation intensity and a preset difference in oscillation intensity, and the increase in the smoothing filter intensity is positively correlated with the difference in oscillation intensity. The reduction in the smoothing filter intensity is determined based on the oscillation intensity offset value and the preset oscillation intensity offset value, and the reduction in the smoothing filter intensity is positively correlated with the oscillation intensity offset value.

6. The spot electricity price forecasting method based on multi-source data fusion according to claim 1, characterized in that, The process of determining whether to adjust the wavelet reconstruction coefficients of each component includes: The corresponding frequency band fidelity coefficient is calculated based on the root mean square of the difference sequence corresponding to each component and the root mean square of the corresponding original sequence, wherein the difference sequence is determined based on the difference between several corresponding original sequences and interpolation sequences at the same time point. The component comparison coefficient is obtained by calculating based on the frequency band fidelity factor and the corresponding preset timeliness factor; Based on the fact that the corresponding component comparison coefficient is less than the corresponding preset component comparison coefficient, the wavelet reconstruction coefficient of the corresponding component is determined and adjusted.

7. The spot electricity price forecasting method based on multi-source data fusion according to claim 6, characterized in that, The process of determining the adjustment coefficients of wavelet reconstruction coefficients based on residual fluctuation values ​​includes: Based on the fact that the residual fluctuation value is greater than or equal to the preset residual fluctuation value, the corresponding adjustment coefficient is reduced. Based on the fact that the residual fluctuation value is less than the preset residual fluctuation value, it is determined to increase the corresponding adjustment coefficient.

8. The spot electricity price forecasting method based on multi-source data fusion according to claim 7, characterized in that, The reduction range of the adjustment coefficient is determined based on the fluctuation difference and the preset fluctuation difference, and the reduction range of the adjustment coefficient is positively correlated with the fluctuation difference; The increase in the adjustment coefficient is determined based on the fluctuation offset value and the preset fluctuation offset value, and the increase in the adjustment coefficient is positively correlated with the fluctuation offset value.

9. The spot electricity price forecasting method based on multi-source data fusion according to claim 1, characterized in that, The process of determining the wavelet transform of the fused time series based on the electricity price quality fluctuation value includes: The signal-to-noise ratio and periodicity of the fused time series are calculated, and the electricity price quality fluctuation value is obtained by calculation. Based on the fact that the electricity price quality fluctuation value is greater than or equal to the preset electricity price quality fluctuation value, it is determined that wavelet transform will be performed on the fused time series sequence.

10. A spot electricity price forecasting system based on multi-source data fusion, used to implement the spot electricity price forecasting method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect multi-source datasets, which include power grid load data, weather forecast data, renewable energy clearing data, regional tie line power flow data, and historical electricity price data. The quality assessment module is used to perform data fusion based on the multi-source dataset and to calculate the electricity price quality fluctuation value based on the generated fused time series. The time-frequency decomposition module is used to determine, based on the electricity price quality fluctuation value, to perform wavelet transform on the fused time series to obtain random fluctuation components, periodic fluctuation components and long-term trend components. The fidelity coefficient generation module is used to extract and time-align each component to obtain the corresponding original sequence and interpolated sequence, and to calculate the corresponding frequency band fidelity coefficient. The weighted fusion module is used to determine the frequency band fidelity factor corresponding to each component based on the frequency band fidelity coefficient, and to determine the component comparison coefficient with the corresponding preset timeliness factor. The coefficient adjustment module is used to determine whether to adjust the wavelet reconstruction coefficients of each component based on the corresponding component comparison coefficients, and to determine the adjustment coefficients of the wavelet reconstruction coefficients based on the residual fluctuation value to obtain the fused feature sequence. The prediction module is used to input the fused feature sequence into the electricity price prediction model to generate an initial prediction sequence; The correction module is used to extract the local outlier degree and local oscillation intensity of the initial prediction sequence, determine whether to correct the initial prediction sequence based on the local outlier degree, and determine the correction amount of the initial prediction sequence based on the local oscillation intensity to generate the final prediction sequence.

Images