Three-dimensional atmosphere live field fusion product generation method based on time sequence analysis

Abstract

The invention discloses a time sequence analysis-based three-dimensional atmosphere live field fusion product generation method, which comprises the following steps of: carrying out preprocessing and space-time matching on multi-source data such as a mode background field and sounding data, training and predicting a 10-minute change trend of the mode field by utilizing a time sequence model, and fusing a predicted variable quantity and observation data after model optimization, so as to obtain a three-dimensional atmosphere live field fusion product. And generating a three-dimensional live field which has a 100-meter resolution ratio and is updated 10 minutes by 10 minutes. According to the method, the bottleneck of multi-source data matching is solved, long-time and short-time prediction is considered, model generalization and physical reasonability are enhanced, fusion precision is improved, and a service application scene is expanded.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of atmospheric science and weather forecasting technology, and particularly relates to a three-dimensional atmospheric real-time field fusion product generation method based on time series analysis. BACKGROUND

[0002] In the field of atmospheric science and weather forecasting, the construction of high-precision three-dimensional atmospheric real-time field is the core basis for short-weather forecasting, disaster weather warning, aviation meteorological support, etc. The traditional numerical weather prediction model generates a prediction field by solving the atmospheric dynamics and thermodynamics equation set, and its output is usually used as a background field for real-time analysis. However, the existing technology faces multiple technical bottlenecks in practical application:

[0003] In terms of data preprocessing and space-time matching, the prediction interval of the conventional numerical model is generally long, while the time resolution of sounding, radar and other observation data is high, and the time scales of the two are not unified. At the same time, the spatial resolution of the model background field is low, and although the ground observation equipment has high vertical or time resolution, the spatial coverage is limited. Traditional interpolation methods cannot effectively align the time and space scales of multi-source data, resulting in poor data consistency and difficulty in meeting the high-precision fusion requirements.

[0004] In terms of time series modeling and prediction, existing methods mostly use a single model architecture. Although traditional LSTM and other models can handle short sequences, they are difficult to capture the long sequence evolution of weather systems. Numerical models based on physical equations are computationally complex and cannot be updated frequently. In addition, existing models have single input features, lack integration of static features such as terrain and surface type, and dynamic statistics of observation data, and have insufficient generalization ability in complex terrain or sudden weather scenarios.

[0005] In terms of data fusion and precision guarantee, existing fusion methods mostly use static weight weighting or simple physical constraints, which cannot dynamically adjust the credibility of different data according to weather conditions and observation density, resulting in poor precision of the fusion product in complex scenarios. Moreover, traditional data-driven methods often ignore atmospheric physical laws, which can easily lead to physical distortion problems; pure physical assimilation methods have high computational cost, making it difficult to balance computational efficiency and physical consistency.

[0006] In terms of application adaptability, the traditional three-dimensional real-time field has low spatial and temporal resolution, which cannot capture small-scale weather features such as thunderstorms and urban heat islands, making it difficult to meet the needs of short-weather forecasting, aviation meteorology, etc. for high-precision and high-frequency data; at the same time, its technical architecture lacks flexibility, making it difficult to efficiently interface with existing meteorological business systems.

[0007] In summary, how to realize the spatio-temporal unified processing of multi-source heterogeneous data, construct a model architecture considering long and short-term prediction, enhance the physical representation and generalization ability of the model, ensure the rationality and precision of prediction, improve the adaptability of the fusion field, and expand business applications, has become a technical problem to be solved in the current meteorological forecasting field. SUMMARY

[0008] The present application aims to provide a three-dimensional atmospheric real-time field fusion product generation method based on time series analysis, which realizes the generation of 100-meter resolution, 10-minute updated three-dimensional atmospheric real-time field by data preprocessing and spatio-temporal matching, time series modeling and optimization, dynamic weighted fusion, improves the short-term and long sequence prediction accuracy, enhances the generalization ability of the model, and meets the demand for high-precision meteorological data in the field of short-term and aviation meteorology.

[0009] To achieve the above purpose, the present application adopts the following technical means:

[0010] A three-dimensional atmospheric real-time field fusion product generation method based on time series analysis, comprising the following steps:

[0011] Data preprocessing and spatio-temporal matching are performed on the model background field, sounding data, microwave radiometer data, wind profile radar data and satellite data to generate a unified spatio-temporal grid data set;

[0012] The time series model is used to train the historical data set to predict the 10-minute change trend of the model background field;

[0013] The model is trained and optimized, and a specific loss function and data enhancement strategy are used to improve the prediction accuracy;

[0014] The predicted change is added to the model background field, and the interpolation results of the multi-source observation data are fused to generate a 100-meter resolution, 10-minute updated three-dimensional atmospheric real-time field.

[0015] As a preferred, the data preprocessing and spatio-temporal matching step comprises:

[0016] Time alignment processing: cubic spline interpolation is used for model data to generate a background field sequence with 10-minute intervals, and optimal interpolation is used for observation data to match to 10-minute timestamps;

[0017] Spatial matching processing: inverse distance weighted interpolation is used for site observation data to 100-meter grid, and super-resolution reconstruction network is used for satellite data to improve the resolution to 100-meter level.

[0018] As a preferred, the input features of the time series model include: sliding statistics of the past 3 hours of model data, the last 1 hour of observation data, and static features of terrain height and surface type.

[0019] Preferably, the selection strategy for the time series model is as follows: use LSTM or ConvLSTM models for short-term prediction to capture local spatiotemporal features; use Transformer + spatiotemporal attention mechanism for long-term sequence modeling to handle long-term sequence dependencies.

[0020] Preferably, the training data for model training is a spatiotemporally aligned sample of historical pattern data and high temporal resolution observation data.

[0021] Preferably, the loss function for model training is a combination of mean squared error and physical constraints, including thermodynamic equilibrium constraints and wind field divergence constraints.

[0022] Preferably, the data augmentation strategy includes: randomly rotating the training data to enhance spatial generalization, and adding Gaussian noise to improve the model's anti-interference ability.

[0023] Preferably, the three-dimensional real-world field fusion generation step includes: superimposing the 10-minute variation predicted by the time series model onto the model background field, combining the spatial interpolation results of multi-source observation data, and optimizing the final field accuracy through dynamic weighted assimilation of model weights and observation weights.

[0024] The present invention has the following beneficial effects:

[0025] 1. Unified spatiotemporal processing of multi-source data to solve the bottleneck of heterogeneous data matching.

[0026] By using cubic spline interpolation and optimal interpolation techniques, 3-hour model data and minute-level observation data are unified to a 10-minute timestamp, eliminating time scale differences. By combining inverse distance weighting and super-resolution reconstruction, single-station observations and low-resolution satellite data are matched to a 100-meter grid, forming a standardized spatiotemporal dataset, laying the foundation for high-precision fusion.

[0027] This processing mechanism solves the core problem of spatiotemporal misalignment between model data and observation data in traditional methods, enabling precise alignment of multi-source data in both time and space dimensions, thereby improving data consistency and availability.

[0028] 2. Hybrid time series model architecture, balancing short-term forecasting and long-term series modeling

[0029] A combined model strategy using LSTM / ConvLSTM and Transformer + spatiotemporal attention is employed.

[0030] The LSTM / ConvLSTM model captures local spatiotemporal features to achieve accurate prediction of rapid changes on a 10-minute scale, making it suitable for short-term evolution scenarios such as severe convective weather.

[0031] The Transformer model uses a spatiotemporal attention mechanism to handle long-distance time dependencies, enhancing its ability to analyze long-term trends in weather systems and adapting to the evolution patterns of different weather processes.

[0032] This architecture avoids the limitations of a single model in terms of time scale adaptability, enabling the model to maintain high accuracy in both short-term forecasts and long-term series analysis, and improving the ability to dynamically track complex weather systems.

[0033] 3. Multi-dimensional feature fusion input enhances the model's physical representation and generalization ability.

[0034] Input features integrate the past 3-hour pattern data sequence, observation data sliding statistics, and topographic height / surface type:

[0035] Dynamic characteristics ensure the capture of temporal evolution trends of atmospheric elements;

[0036] Observational statistical features enhance the contribution of real-time data and improve the model's ability to respond to sudden weather events;

[0037] Static features reflect the influence of the underlying surface, avoiding the prediction bias of traditional models in complex terrains such as mountains and cities.

[0038] Multi-dimensional feature fusion enables the model to learn atmospheric dynamics and geographical environmental influences simultaneously, significantly improving its generalization ability across different geographical regions.

[0039] 4. Combining physical constraints with data augmentation ensures the rationality and robustness of predictions.

[0040] The loss function incorporates physical constraints such as thermodynamic equilibrium and wind field divergence, deeply coupling data-driven prediction results with atmospheric physical laws. This avoids physical distortions such as abnormal vertical temperature distribution and non-closed wind fields that may occur with pure machine learning methods, ensuring the physical consistency of the three-dimensional field.

[0041] Data augmentation strategies enhance the model's resilience to complex observation conditions and reduce abnormal predictions in practical applications by simulating the spatial diversity and noise interference of real observation environments.

[0042] 5. Dynamic weighted assimilation mechanism to improve the accuracy and adaptability of the fusion field.

[0043] By overlaying the model's predicted 10-minute variations onto the model's background field and dynamically adjusting the model and observation weights based on observation density, the model is fused.

[0044] In densely observed areas, enhance the contribution of real-time observation data to improve the accuracy of characterization of local fine features;

[0045] In sparse observation regions, the physical constraints of the model background field are relied upon to ensure the rationality of the large-scale field.

[0046] This mechanism addresses the shortcomings of traditional fusion methods where model and observation weights are fixed, ensuring that the generated 100-meter / 10-minute 3D field maintains consistent accuracy across the entire region, making it particularly suitable for complex scenarios with significant differences in observation conditions.

[0047] 6. High-frequency, high-precision output expands business application scenarios.

[0048] The generated 3D real-time field achieves a spatial resolution of 100 meters and updates every 10 minutes, improving the spatiotemporal resolution by 18 times and 10 times respectively compared to traditional models. It can capture small- and medium-scale weather features such as thunderstorms, wind shear, and urban heat islands, providing core data support for fields with stringent requirements for spatiotemporal accuracy, such as short-term forecasting, aviation meteorology, and urban climate simulation.

[0049] The modular technical architecture supports seamless integration with existing meteorological business systems, meets real-time update requirements, and significantly improves the efficiency and adaptability of business applications. Detailed Implementation

[0050] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1

[0052] A method for generating a three-dimensional atmospheric field fusion product based on time series analysis includes the following steps:

[0053] 1. Data preprocessing and spatiotemporal matching

[0054] Input data:

[0055] Model background field: spatiotemporal resolution 3 hours / 1km, including temperature, humidity, and wind field;

[0056] Radiosonde data: 12-hour spatiotemporal resolution / high vertical resolution, including temperature, humidity, and wind;

[0057] Microwave radiometer: 10-minute spatiotemporal resolution / low vertical resolution, including temperature and humidity;

[0058] Wind profiler radar: spatiotemporal resolution of 6 minutes / 100m vertical resolution, including wind field;

[0059] Satellite data: used to supplement large-scale atmospheric information.

[0060] Time alignment processing:

[0061] Cubic spline interpolation was used to generate background field sequences with 10-minute intervals from the model data;

[0062] The observation data is matched to a 10-minute timestamp using optimal interpolation (OI) to ensure time synchronization of multi-source data.

[0063] Spatial matching processing:

[0064] For observation data from stations such as radiosonde, microwave radiometer, and wind profiler radar, inverse range weighted (IDW) interpolation was used to a 100m grid.

[0065] For satellite data, a super-resolution reconstruction network (SRCNN) is used to improve the spatial resolution to the 100m level.

[0066] 2. Time Series Modeling and Dynamic Forecasting

[0067] Model input: Model data (temperature, humidity, wind field) for the past 3 hours (18 10-minute steps), and moving statistics (mean, trend) of the most recent hour's observations.

[0068] Static features: terrain height and surface type (One-Hot encoding), used to enhance spatial context information.

[0069] Model selection:

[0070] Short-term prediction: Employ LSTM or ConvLSTM models to capture local spatiotemporal features;

[0071] Long sequence modeling: A Transformer + spatiotemporal attention mechanism is used to handle the dependencies of long sequence.

[0072] 3. Model Training and Optimization

[0073] Training data: Spatiotemporally aligned samples of historical model data (3-hour resolution) and high temporal resolution observation data (wind profiler radar, microwave radiometer).

[0074] Loss function: Mean squared error (MSE) + physical constraint terms (such as thermodynamic balance and wind field divergence constraints) to ensure that the prediction results conform to atmospheric physical laws.

[0075] Data augmentation: Spatial generalization is enhanced by random rotation (90° / 180° / 270°), and Gaussian noise (σ=0.1) is added to improve the model's robustness to interference.

[0076] 4. Generation of 3D Real-Time Scene Fusion

[0077] The 10-minute changes (ΔT, ΔRH, ΔU, ΔV) predicted by the time series model are superimposed on the model background field to generate a three-dimensional field at the predicted time.

[0078] Spatial interpolation results from multi-source observation data are integrated, and the accuracy of the final field is optimized through weighted assimilation (model weight + observation weight).

[0079] Example of a model structure (using ConvLSTM as an example)

[0080] import tensorflow as tf

[0081] from tf.keras.layers import ConvLSTM2D, BatchNormalization

[0082] model = tf.keras.Sequential([

[0083] # Encoder (Spatiotemporal Feature Extraction)

[0084] ConvLSTM2D(filters=64, kernel_size=(3,3),

[0085] input_shape=(18, 100, 100, 4), # 18 10-minute time steps, 100x100 grid, 4 variables

[0086] return_sequences=True)

[0087] BatchNormalization(),

[0088] # Bottleneck layer (information compression)

[0089] ConvLSTM2D(filters=128, kernel_size=(3,3),

[0090] padding='same', return_sequences=False),

[0091] # Decoder (predicting changes)

[0092] Conv2D(filters=4, kernel_size=(1,1), activation='linear') # Output ΔT, ΔRH, ΔU, ΔV

[0093] ]).

[0094] Example 2

[0095] Real-time analysis of three-dimensional temperature field in urban areas

[0096] Application scenarios

[0097] The requirement is for short-term temperature forecasting and urban heat island effect monitoring in a megacity (approximately 1000 km², with a terrain mainly consisting of plains and dense building clusters), with a temperature field updated every 10 minutes at a resolution of 100 meters.

[0098] Technical process and parameter configuration

[0099] 1. Data Preprocessing

[0100] Input data:

[0101] Background field of the mode: 3 hours / 1km resolution, including 2m temperature;

[0102] Observational data:

[0103] Microwave radiometer: 10 minutes / time, single-station vertical temperature profile (low vertical resolution).

[0104] Ground-based automatic weather stations: 2m temperature readings at discrete stations (50 stations within the city, spaced approximately 2km apart) every 1 minute;

[0105] Satellite data: MODIS surface temperature product (500m resolution).

[0106] Spatiotemporal matching:

[0107] Time alignment: Cubic spline interpolation is used to generate 10-minute sequences for model data, and observation data is matched to 10-minute timestamps through optimal interpolation;

[0108] Spatial matching: Automatic station data is interpolated to a 100m grid (distance power exponent p=2) using inverse distance weighted (IDW) method, and satellite data is super-resolutiond to 100m using SRCNN.

[0109] 2. Time series modeling

[0110] Model selection: ConvLSTM model (suitable for predicting short-term local changes in urban areas);

[0111] Input features:

[0112] Dynamic features: Pattern temperature data (100×100 grid) over the past 3 hours (18 steps).

[0113] Observational statistics: the moving average and rate of change of temperature from automatic weather stations over the past hour;

[0114] Static features: terrain height (100m resolution), land surface type (One-Hot encoding of urban / green space / water body).

[0115] 3. Model Training and Fusion

[0116] Training data: Model data and observation data of the city in the summer of 2024 (June-August), totaling 30 days, generating 1440 training samples;

[0117] Loss function: MSE + vertical temperature lapse rate constraint (λ=0.1);

[0118] Fusion generation: The model-predicted 10-minute temperature change is superimposed onto the model field, and combined with automatic weather station and satellite interpolation data (model weight α=0.6) to generate the final temperature field.

[0119] Implementation effect

[0120] Compared with the traditional 3-hour model direct interpolation, the root mean square error of temperature (RMSE) decreased from 1.2℃ to 0.7℃, especially during the peak period of the urban heat island effect in the afternoon (14:00-16:00), the accuracy of temperature gradient characterization in building cluster areas was improved by 40%;

[0121] Successfully captured the daily temperature variation differences between the city center and the suburbs, providing refined data support for urban energy dispatch and high temperature early warning.

[0122] Example 3

[0123] Short-term wind field forecast for airport area

[0124] Application scenarios

[0125] A wind shear warning for an international airport (runway length 3km, surrounding terrain includes hills) requires a 3D wind field (U / V / W components) with a vertical resolution of 100 meters and updated every 10 minutes.

[0126] Technical process and parameter configuration

[0127] 1. Data Preprocessing

[0128] Input data:

[0129] Model background field: 3 hours / 1km resolution, including 10m wind field;

[0130] Observational data:

[0131] Wind profiler radar: 6 minutes / time, vertical resolution 100m, single-station wind field profile (wind speed, wind direction).

[0132] Radiosonde data: 12 hours / time, high vertical resolution wind field (sounding station near the airport).

[0133] Spatiotemporal matching:

[0134] Time alignment: The model data is interpolated to 10 minutes using cubic splines, and the radar data is matched to the timestamp using optimal interpolation;

[0135] Spatial matching: Wind profiler radar data is interpolated to a 100m grid of 30km×30km around the airport by inverse range weighting.

[0136] 2. Time series modeling

[0137] Model selection: Transformer + spatiotemporal attention model (for handling long-sequence evolution of airport wind fields, such as low-level jet streams).

[0138] Input features:

[0139] Dynamic characteristics: Model wind field data over the past 6 hours (36 steps);

[0140] Observational statistics: Vertical shear statistics of radar wind field in the past hour;

[0141] Static features: airport terrain elevation, runway orientation (One-Hot encoding).

[0142] Attention mechanism: Spatial attention focuses on the runway area, while temporal attention captures the periodic changes in the wind field.

[0143] 3. Model Training and Fusion

[0144] Training data: Model data and radar data of the airport in the spring of 2024 (March-May), totaling 90 days, generating 5184 training samples;

[0145] Loss function: MSE + wind field divergence constraint (λ=0.2);

[0146] Fusion generation: The predicted wind field changes are superimposed on the model field and then dynamically weighted and fused with radar interpolation data (model weight α=0.5, and α is reduced to 0.3 in densely observed areas).

[0147] Implementation effect

[0148] The lead time for identifying low-level wind shear has been extended from 15 minutes to 30 minutes using traditional methods, the wind speed RMSE has decreased from 1.8 m / s to 1.0 m / s, and the wind direction error has decreased from 8° to 5°.

[0149] Successfully capturing local wind field disturbances caused by hilly terrain provides crucial data for wind shear warning during aircraft takeoff and landing, enhancing flight safety capabilities.

[0150] Example 4

[0151] Refined analysis of three-dimensional humidity field in mountainous areas

[0152] Application scenarios

[0153] For heavy rainfall monitoring in a mountainous watershed (approximately 500 km² with an elevation difference of 1500 m), the requirement is a relative humidity field with a vertical resolution of 100 meters and updated every 10 minutes, for flash flood early warning.

[0154] Technical process and parameter configuration

[0155] 1. Data Preprocessing

[0156] Input data:

[0157] Background field of the mode: 3 hours / 1km resolution, including relative humidity;

[0158] Observational data:

[0159] Microwave radiometer: 10 minutes / time, single-station humidity vertical profile (1 unit each at the foot of the mountain and the middle of the mountain).

[0160] Radiosonde data: 12 hours / time, humidity profile of stations near mountainous areas;

[0161] Satellite data: Remote sensing data of water vapor channels (1km resolution).

[0162] Spatiotemporal matching:

[0163] Satellite data is super-resolutiond to 100m using SRCNN, and vertical interpolation is performed in conjunction with terrain height to correct humidity deviations caused by altitude differences in mountainous areas.

[0164] 2. Time series modeling

[0165] Model selection: LSTM model (mountain weather systems exhibit significant short-term variations, suitable for short-term forecasting);

[0166] Input features:

[0167] Dynamic characteristics: Pattern humidity data over the past 3 hours (18 steps);

[0168] Static features: terrain height and mountain orientation (One-Hot encoding) to enhance the influence of terrain on humidity distribution;

[0169] Model optimization: Add Gaussian noise (σ=0.1) to improve the model's robustness to complex observation conditions in mountainous areas.

[0170] 3. Model Training and Fusion

[0171] Loss function: MSE + thermodynamic equilibrium constraint (coupling relationship between saturated water vapor pressure and temperature, λ=0.15);

[0172] Fusion strategy: In sparsely observed areas in mountainous regions, increase model weight (α=0.8) to rely on the physical constraints of the model; in areas covered by microwave radiometers, decrease model weight (α=0.4) to enhance the contribution of observational data.

[0173] Implementation effect

[0174] The relative humidity RMSE was reduced from 8% to 5% by traditional methods, and the accuracy of humidity gradient characterization was improved by 35% in mountainous areas, especially on the windward and leeward slopes, effectively identifying the phenomenon of sudden increase in humidity before heavy precipitation.

[0175] It provides refined humidity evolution data for flash flood early warning systems, increasing the lead time for precipitation forecasts by 20 minutes and improving the timeliness of disaster prevention in mountainous areas.

[0176] The examples provided in this invention are not intended to limit the implementation. Those skilled in the art will recognize that various variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations, and any obvious variations or modifications derived therefrom are still within the scope of this invention.

Claims

1. A method for generating a three-dimensional atmospheric field fusion product based on time series analysis, characterized in that, Includes the following steps: Data preprocessing and spatiotemporal matching are performed on the background field of the model, radiosonde data, microwave radiometer data, wind profiler radar data, and satellite data to generate a dataset with a unified spatiotemporal grid. Using a time series model trained on a historical dataset, the 10-minute trend of pattern background field changes is predicted; The model is trained and optimized, and specific loss functions and data augmentation strategies are used to improve prediction accuracy. The predicted changes are superimposed on the model background field and the interpolation results of multi-source observation data are fused to generate a three-dimensional atmospheric field with a resolution of 100 meters and updated every 10 minutes.

2. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The data preprocessing and spatiotemporal matching steps include: Time alignment processing: Cubic spline interpolation is used to generate a background field sequence with 10-minute intervals for the model data, and optimal interpolation is used to match the observation data to the 10-minute timestamp; Spatial matching processing: Inverse distance weighted interpolation is used to interpolate the station observation data to a 100-meter grid, and super-resolution reconstruction network is used to improve the resolution of satellite data to the 100-meter level.

3. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The input features of the time series model include: the model data of the past 3 hours, the sliding statistics of the observation data of the most recent 1 hour, and the static features of terrain height and land surface type.

4. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The selection strategy for the time series model is as follows: use LSTM or ConvLSTM models for short-term prediction to capture local spatiotemporal features; use Transformer + spatiotemporal attention mechanism for long-term sequence modeling to handle long-term sequence dependencies.

5. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The training data for the model training consists of spatiotemporally aligned samples of historical pattern data and high temporal resolution observation data.

6. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The loss function for training the model is a combination of mean squared error and physical constraints, including thermodynamic equilibrium constraints and wind field divergence constraints.

7. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The data augmentation strategies include: randomly rotating the training data to enhance spatial generalization and adding Gaussian noise to improve the model's anti-interference ability.

8. The method for generating a three-dimensional atmospheric field fusion product based on time series analysis according to claim 1, characterized in that, The three-dimensional real-world field fusion generation steps include: superimposing the 10-minute variation predicted by the time series model onto the model background field, combining the spatial interpolation results of multi-source observation data, and optimizing the final field accuracy through dynamic weighted assimilation of model weights and observation weights.