Power load prediction method and device based on solar term temperature

Abstract

The invention relates to a solar term temperature-based power load prediction method, which comprises the following steps of: obtaining 24 solar term reference temperatures and long-term trends of solar term temperatures, and correcting the long-term trends of the solar term temperatures; generating a continuous basic seasonal temperature curve; generating a high-resolution final simulation temperature sequence; and inputting the high-resolution final simulation temperature sequence into a power load prediction model, and performing power load prediction. Through throttle window smoothing, Gaussian process regression kernel function physical constraint and physical amplitude limiting random noise design, high-frequency random noise and chaos disturbance in original observation temperature are effectively stripped, the sensitivity of a power load prediction model to local atypical weather fluctuation is remarkably reduced, overfitting is relieved fundamentally, and the power load prediction method has the advantages of being high in robustness and high in reliability. The generalization capability and the prediction stability of the power load prediction model are greatly improved; and main physical driving factors of temperature change are subjected to decoupling and explicit modeling, so that the interpretability of the model is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load forecasting, in particular to a power load forecasting method and device based on solstice temperature. BACKGROUND

[0002] Accurate prediction of power load is the core link to ensure the safe and stable operation of power grid and optimize economic dispatch. Existing prediction methods mainly include traditional statistical methods, machine learning methods and deep learning methods. Traditional statistical methods such as time series analysis and regression analysis are difficult to handle nonlinear relationships. Machine learning methods such as random forest and support vector machine require high data quality. Deep learning methods such as LSTM and GRU recurrent neural network can capture time series dependence.

[0003] There are two major bottleneck problems in existing prediction methods: first, the lack of physical laws, existing models cannot inherently understand the underlying physical mechanism of the annual periodic variation of solar radiation determined by the orbit parameters of the earth; second, local disturbance sensitivity, the model is too sensitive to micro and high frequency noise caused by chaotic atmospheric activities such as turbulence and advection, leading to overfitting.

[0004] In addition, traditional observation temperature methods cannot fundamentally solve the above problems because they fail to effectively remove noise and embed physical prior constraints. Therefore, there is an urgent need for a new load forecasting method that can integrate physical laws and data-driven methods. SUMMARY

[0005] To solve the problem of lack of astronomical background signal in traditional observation temperature and significant and unpredictable fluctuations in short-term measurement results, the primary purpose of the present application is to provide a physical tool for converting the traditional Chinese 24 solar terms system into a defined astronomical anchor point, significantly reducing the sensitivity of the power load forecasting model to local atypical weather fluctuations, and greatly improving the generalization ability and prediction stability of the power load forecasting model. A power load forecasting method based on solstice temperature.

[0006] To achieve the above purpose, the present application adopts the following technical solution: a power load forecasting method based on solstice temperature, which comprises the following sequential steps:

[0007] (1) Obtain the 24 solar term reference temperature and the long-term trend of the solstice temperature, correct the long-term trend of the solstice temperature, and obtain the long-term warming trend item ΔT trend (y) and the smoothed reference temperature value T s * of 24 solar terms;

[0008] (2) According to the smoothed reference temperature value T s* and Gaussian process regression model to generate continuous underlying seasonal temperature curve T GPR(d);

[0009] (3) according to long-term warming trend term ΔT trend (y) and continuous underlying seasonal temperature curve T GPR(d), generate high-resolution final simulated temperature sequence T final(t);

[0010] (4) input high-resolution final simulated temperature sequence T final(t) into power load prediction model to perform power load prediction.

[0011] Step (1) specifically comprises the following steps:

[0012] (1a) use Chinese traditional 24 solar terms as astronomical anchor points to calibrate specific positions of Earth's orbit, each solar term corresponds to a specific solar declination δ, and obtain 24 solar term reference temperatures;

[0013] (1b) collect historical temperature data of solar term temperature in a long time span, i.e. long-term trend;

[0014] (1c) for each solar term point, correct the historical temperature data using a correction formula to obtain long-term warming trend term ΔT trend (y):

[0015] ΔT trend (y) = β o (y-y ref ) + f osc (y);

[0016] Wherein, y is the year, y ref is the reference base year, β o is the annual warming rate, f osc (y) represents the contribution of the main climate mode to the annual anomaly;

[0017] (1d) set an extended time window before and after each solar term point of the 24 solar term reference temperature, and perform weighted average or local polynomial fitting on the historical temperature data in the extended time window to obtain smoothed reference temperature values T s * of the 24 solar terms.

[0018] Step (2) specifically comprises the following steps:

[0019] (2a) use Gaussian process regression model to generate continuous underlying seasonal temperature curve T GPR(d) using smoothed reference temperature values T s * of the 24 solar terms as training data;

[0020] (2b) construct a triple composite kernel function k(d,d') to encode the physical law of temperature change:

[0021] k(d,d') = k period (d,d') x k se (d,d') + k noise (d,d');

[0022] where k period is a periodic kernel; k se is a square exponential kernel; k noise is a white noise kernel; d, d' are both date serial numbers in a year.

[0023] Step (3) specifically comprises the following steps:

[0024] (3a) take the continuous base seasonal temperature curve T GPR(d) as the main trunk;

[0025] (3b) on the basis of the main trunk, superimpose the intra-daily periodic fluctuation to introduce the deterministic daily variation component ΔT d iur (h):

[0026] ΔT d iur (h) = A(d) · sin(2π(h-h0) / 24);

[0027] where h is the hour, h0 is the time when the lowest temperature occurs, and A(d) is the daily temperature difference varying with the season;

[0028] (3c) further superimpose the physically limited random noise to introduce the physically constrained random noise term ζ(d, h):

[0029] ζ(d, h) ~ N(0, σ season 2 (d)) s.t. |ζ|≤Γ;

[0030] where σ season 2 (d) is the noise standard deviation, σ season 2 (d) varies with the season, and Γ represents the limit of possible weather disturbance in the region to avoid generating extreme jumps that violate physical common sense;

[0031] (3d) further superimpose the long-term warming trend to introduce the long-term warming trend term ΔT trend (y);

[0032] (3e) finally, generate the high-resolution final simulated temperature sequence T final(t):

[0033] T final(t) = T GPR(d) + ΔT d iur (h) + ζ(d, h) + ΔT trend (y).

[0034] Step (4) specifically comprises the following steps:

[0035] (4a) inputting the final simulation temperature sequence T_final(t) and other related features into a power load prediction model; the other related features include historical load, humidity and electricity price, and the other related features are obtained from the NSW subset of the load and meteorological data set published by AEMO; the power load prediction model adopts any one of bidirectional long short-term memory network BiLSTM, LSTM, GRU, CNN-LSTM and BiGRU model;

[0036] (4b) training the power load prediction model, so that the power load prediction model learns the mapping relationship between the load and the temperature field, thereby improving the prediction accuracy;

[0037] (4c) using the trained model to predict the future power load.

[0038] Another object of the present application is to provide an electronic device comprising:

[0039] a processor; and

[0040] a memory, in which computer program instructions are stored, the computer program instructions, when executed by the processor, cause the processor to perform the power load prediction method based on the weather temperature as described above.

[0041] The present application also provides a computer readable storage medium having computer program instructions stored thereon, the computer program instructions, when executed by a processor, cause the processor to perform the power load prediction method based on the weather temperature as described above.

[0042] As can be seen from the above technical solution, the beneficial effects of this invention are as follows: First, enhanced physical mechanism: This invention creatively transforms the traditional Chinese 24 solar terms system into a physical tool for defining astronomical anchor points, providing a deterministic framework for core temperature changes determined by the Earth's revolution, thus solving the fundamental problem of the lack of physical mechanisms in data-driven models; Second, improved noise removal and robustness: This invention effectively removes high-frequency random noise and chaotic disturbances from the original observed temperature through solar term window smoothing, kernel function physical constraints of Gaussian process regression, and physical amplitude-limited random noise design, significantly reducing the sensitivity of the power load forecasting model to local atypical weather fluctuations, fundamentally alleviating overfitting, and greatly improving the generalization ability and prediction stability of the power load forecasting model; Third, significantly improved prediction accuracy: Experiments have shown that… Applying the temperature field generated by this invention for load forecasting reduces the mean absolute percentage error by 39.16% and the root mean square error by 36.57% compared to traditional methods that directly use raw observed temperatures. Even during peak load periods such as the summer solstice and critical periods of severe weather fluctuations, the improvement in forecast accuracy and robustness is more significant. Fourth, the model has strong universality: the temperature field generated by this invention, as a high-quality, high-physical-fidelity input feature, has had its performance gains consistently verified across various mainstream load forecasting model architectures, indicating its broad applicability. Fifth, interpretability is improved: the four-dimensional temperature synthesis mechanism, namely the superposition of reference temperature, daily cycle, amplitude-limiting noise, and long-term trend, decouples and explicitly models the main physical driving factors of temperature changes, enhancing the interpretability of the model. Attached Figure Description

[0043] Figure 1 This is a flowchart of the method of the present invention;

[0044] Figure 2 This is a schematic diagram of the prediction results based on the BiLSTM network on the Australian dataset of this invention. Detailed Implementation

[0045] like Figure 1 As shown, a power load forecasting method based on seasonal temperature includes the following steps in sequence:

[0046] (1) Obtain the reference temperature and long-term trend of the solar term temperature for the 24 solar terms, correct the long-term trend of the solar term temperature, and obtain the long-term warming trend term ΔT. trend (y) and the smoothed reference temperature value T of the 24 solar terms s *;

[0047] (2) Based on the smoothed reference temperature value T of the 24 solar terms s * and Gaussian process regression model to generate continuous basic seasonal temperature curve T_GPR(d);

[0048] (3) Based on the long-term warming trend term ΔT trend (y) and continuous basic seasonal temperature curves T_GPR(d) are used to generate a high-resolution final simulated temperature sequence T_final(t);

[0049] (4) Input the high-resolution final simulated temperature sequence T_final(t) into the power load prediction model to perform power load prediction.

[0050] Step (1) specifically includes the following steps:

[0051] (1a) Using the traditional Chinese 24 solar terms as astronomical anchor points to mark specific positions in the Earth's orbit, each solar term corresponds to a specific solar declination δ, and the 24 solar term reference temperature is obtained.

[0052] (1b) Collect historical temperature data of solar terms over a long period of time, i.e., long-term trends;

[0053] (1c) For each solar term point, the historical temperature data is corrected using the correction formula to obtain the long-term warming trend term ΔT. trend (y):

[0054] ΔT trend (y)=β o (yy ref )+f osc (y);

[0055] Where y is the year, y ref For reference to the base year, β o f is the interannual warming rate. osc (y) represents the contribution of major climate modes to interannual anomalies. These major climate modes refer to large-scale atmospheric and oceanic oscillation models that influence global or regional climate variability. These modes lead to interannual variations in climate elements such as temperature and precipitation. Therefore, when analyzing long-term warming trends, it is necessary to separate their influence from the data to obtain a more accurate trend term. Major climate modes typically include the following common models: El Niño-Southern Oscillation, North Atlantic Oscillation, Pacific Decadal Oscillation, Indian Ocean Dipole, and Antarctic Oscillation.

[0056] Step (1c) is used to eliminate long-term non-astronomical periodic disturbances and obtain the solar term reference temperature that reflects the pure astronomical-seasonal cycle.

[0057] (1d) Set extended time windows before and after each solar term point of the 24 solar term reference temperature, and perform weighted averaging or local polynomial fitting on the historical temperature data within the extended time window to obtain the smoothed reference temperature value T for the 24 solar terms. s *

[0058] Step (2) specifically includes the following steps:

[0059] (2a) Using a Gaussian process regression model, with the smoothed baseline temperature T of the 24 solar terms as the reference value. s * As training data, a continuous baseline seasonal temperature curve T_GPR(d) is generated;

[0060] (2b) Construct a triple composite kernel function k(d,d') to encode the physical laws governing temperature changes:

[0061] k(d,d')=k period (d,d')×k se (d,d')+k noise (d,d');

[0062] Where, k period For periodic kernels; k se For the squared exponent kernel; k noise is a white noise kernel; d and d' are both ordinal numbers of dates in a year.

[0063] Step (3) specifically includes the following steps:

[0064] (3a) The continuous basic seasonal temperature curve T_GPR(d) is used as the backbone;

[0065] (3b) Based on the main trunk, superimpose intraday periodic fluctuations and introduce a deterministic daily variation component ΔT_d. iur (h):

[0066] ΔT_d iur (h)=A(d)·sin(2π(h-h0) / 24);

[0067] Where h is the hour; h0 is the time when the lowest temperature occurs; A(d) is the daily temperature difference that varies with the seasons, with the peak usually occurring around 14:00, reflecting the delayed effect of heat accumulation;

[0068] (3c) Then, physical limiting random noise is superimposed to introduce a physically constrained random noise term ζ(d,h):

[0069] ζ(d,h)~N(0,σ season 2 (d))st|ζ|≤Γ;

[0070] Where, σ season 2 (d) represents the noise standard deviation, σ season 2 (d) Varies with the seasons; Γ represents the limit of possible weather disturbances in the region's physics, avoiding the generation of extreme jumps that violate common sense in physics;

[0071] (3d) The long-term warming trend is then superimposed, introducing the long-term warming trend term ΔT. trend (y);

[0072] (3e) Finally, the high-resolution final simulated temperature sequence T_final(t) is generated:

[0073] T_final(t)=T_GPR(d)+ΔT_d iur (h)+ζ(d,h)+ΔT trend (y).

[0074] Step (4) specifically includes the following steps:

[0075] (4a) Input the final simulated temperature sequence T_final(t) along with other relevant features into the power load prediction model; the other relevant features include historical load, humidity and electricity price, and the other relevant features are obtained from the NSW subset of the load and meteorological datasets publicly released by AEMO; the power load prediction model adopts any one of the following models: Bi-directional long short-term memory network BiLSTM, LSTM, GRU, CNN-LSTM, and BiGRU.

[0076] (4b) Train the power load forecasting model so that it learns the mapping relationship between load and temperature field, thereby improving the forecasting accuracy;

[0077] (4c) Use the trained model to predict future electricity load.

[0078] like Figure 2 As shown, tests on a dataset from a region in Australia demonstrate that using the temperature field generated by this invention—the high-resolution final simulated temperature sequence T_final(t)—for load prediction, compared to the method using the original temperature, results in: a decrease in the overall mean absolute percentage error (MAPE) from 1.20% to 0.73%, a reduction of 39.16%; and a decrease in the root mean square error (RMSE) from 122.9675 to 77.9975, a reduction of 36.57%. On the peak load day, the summer solstice, the MAPE decreases from 1.50% to 0.81%, a reduction of 46.00%. Even on different model architectures, such as LSTM, GRU, CNN-LSTM, and BiGRU, the temperature field generated by this invention consistently delivers significant accuracy improvements.

[0079] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for power load forecasting based on solstice temperature, characterized in that: The method comprises the following steps in sequence: (1) Obtain the long-term trend of the 24 solar term reference temperature and solar term temperature, correct the long-term trend of the solar term temperature, and obtain the long-term warming trend item ΔT trend (y) and the smoothed reference temperature value T of the 24 solar terms s *; (2) Smooth the base temperature value T according to the 24 solar terms s * and a Gaussian process regression model to generate a continuous base seasonal temperature curve T GPR(d); (3) according to the long-term warming trend term ΔT trend (y) and the continuous underlying seasonal temperature curve T GPR(d), a high-resolution final simulated temperature sequence T final(t) is generated; (4) inputting the high-resolution final simulated temperature sequence T_final(t) into a power load prediction model to perform power load prediction. 2.The power load forecasting method based on the solstice temperature according to claim 1, characterized in that: Step (1) specifically comprises the following steps: (1a) taking the Chinese traditional 24 solar terms as astronomical anchor points for calibrating specific positions of the Earth's orbit, each solar term corresponding to a specific solar declination δ, to obtain 24 solar term reference temperatures; (1b) collecting historical temperature data of solar term temperatures over a long time span, i.e., long-term trends; (1c) For each solstice point, correct the historical temperature data using the correction formula to obtain the long-term warming trend term ΔT trend (y): ΔT trend (y) = β o (y - y ref ) + f osc (y) where y is the year, y ref is the reference base year, β o is the interannual warming rate, f osc (y) represents the contribution of the leading climate mode to the interannual anomaly. (1d) Set an extended time window before and after each solstice point of the 24 solar term reference temperature, and obtain the smoothed reference temperature value T of the 24 solar term by weighted average or local polynomial fitting of the historical temperature data in the extended time window s * 3.The power load forecasting method based on the solstice temperature according to claim 1, characterized in that: Step (2) specifically comprises the following steps: (2a) using a Gaussian process regression model, to smooth the benchmark temperature values T s * as training data, generate a continuous base seasonal temperature curve T GPR(d); (2b) constructing a triple composite kernel function k(d,d') for encoding the physical law of temperature change: k(d,d') = k period (d,d') x k se (d,d') + k noise (d,d'); where k period is the periodic kernel; k se is the squared exponential kernel; k noise is the white noise kernel; d, d' are both the date ordinal in a year. 4.The method according to claim 1, wherein: Step (3) specifically comprises the following steps: (3a) taking the continuous basic seasonal temperature curve T_GPR(d) as the backbone; (3b) On top of the backbone, superimpose the intra-day periodic fluctuations, introduce the deterministic daily variation component ΔΤ_d iur (h): ΔT_d iur (h) = A(d) - sin(2π(h - h0) / 24); where h is the hour, h0 is the time when the lowest temperature occurs, and A(d) is the daily temperature difference varying with the season; (3c) further superimposing a physically limited random noise to introduce a random noise term ζ(d,h) subject to physical constraints: ζ(d, h) ~ N(0, σ season 2 (d) s.t. |ζ|≤Γ; where σ season 2 (d) is the noise standard deviation, σ season 2 (d) varies with season; Γ represents the limit of the regional physical possible weather disturbance to avoid generating extreme jumps that violate physical common sense; (3d) superimposing a long-term warming trend, introducing a long-term warming trend term ΔT trend (y); (3e) finally, generating a high-resolution final simulated temperature sequence T_final(t): T_final(t) = T_GPR(d) + ΔT_d iur (h) + ζ(d, h) + ΔT trend (y). 5.The power load forecasting method based on the solstice temperature according to claim 1, characterized in that: Step (4) specifically comprises the following steps: (4a) inputting the final simulated temperature sequence T_final(t) into a power load prediction model together with other relevant features; the other relevant features include historical load, humidity, and electricity price, and the other relevant features are obtained from the NSW subset of the load and weather dataset published by AEMO; the power load prediction model adopts any one of bidirectional long short-term memory network BiLSTM, LSTM, GRU, CNN-LSTM, and BiGRU model; (4b) training the power load prediction model to enable the power load prediction model to learn the mapping relationship between load and temperature field, thereby improving the prediction accuracy; (4c) using the trained model to perform future power load prediction.

6. An electronic device comprising: a processor; and a memory having computer program instructions stored therein, the computer program instructions causing the processor to execute the solar term temperature-based power load prediction method according to any one of claims 1-5 when executed by the processor.

7. A computer-readable storage medium having computer program instructions stored thereon, the computer program instructions causing the processor to execute the solar term temperature-based power load prediction method according to any one of claims 1-5 when executed by the processor.

Images