The dataset used in this study originates from the Household Data package provided by the Open Power System Data (OPSD) platform and was acquired through public download from the official OPSD website. Open Power System Data is an open-access data platform dedicated to electricity system research, providing curated, validated, processed, and documented datasets collected from publicly available sources. This study directly downloaded and utilized the publicly available dataset from the official OPSD website. No modifications were made to the original data collection process; however, the downloaded data were further processed and curated to meet the requirements of subsequent analysis and modeling.

The dataset was processed based on the Household Data package released by the Open Power System Data (OPSD) platform. All preprocessing procedures were implemented using Python and the open-source data analysis libraries Pandas and NumPy. First, the raw data were imported and standardized with respect to timestamp formatting. Valid UTC timestamps were extracted using regular expressions and converted into a unified datetime format, while timezone information was removed to ensure consistency across all records. Auxiliary fields not required for subsequent analysis, such as daylight-saving-time indicators and interpolation flags, were removed. Next, a data cleaning procedure was performed. Missing-value identifiers in the original files were replaced with standard NaN values. Columns containing only missing values, records with no valid observations except timestamps, all-zero columns, and columns with constant values were removed to eliminate redundant and non-informative features. Subsequently, each residential household dataset was processed individually. Variable-name prefixes associated with different households were removed to establish a consistent naming convention across datasets. Data samples were further filtered according to the selected study period. For key variables related to electricity consumption, photovoltaic generation, and grid interaction, records containing missing target values were excluded to ensure data completeness and reliability. Finally, the cleaned household load, photovoltaic generation, and appliance-level electricity consumption data were organized into a standardized time-series dataset suitable for load forecasting, demand response, energy management, and reinforcement learning-based energy scheduling studies.

Systematic quality-control procedures were applied during data acquisition, cleaning, and integration to ensure the completeness, consistency, and usability of the dataset. The original dataset contained 38,454 time-series records and 71 data fields, covering the period from December 11, 2014, to May 1, 2019, with an hourly temporal resolution. After missing-value inspection, invalid-feature removal, and data standardization, the final dataset retained 38,454 valid records and 69 data fields, with two non-informative features removed. The dataset includes 11 categories of users, consisting of three industrial users (industrial1–industrial3), two public users (public1–public2), and six residential users (residential1–residential6). The residential-user data were further organized into six independent household datasets corresponding to residential1 through residential6. All datasets were successfully generated, with standardized variable naming conventions, consistent timestamp formats, and no duplicate temporal records. Completeness checks were performed on key variables. The retained residential datasets contain valid observations for electricity import from the grid (grid_import), photovoltaic generation (pv), and appliance-level loads such as dishwashers, freezers, washing machines, and heat pumps. Records with missing values in critical target variables were excluded, ensuring that the final datasets contain no missing values in key variables. For example, the residential1 dataset contains 8,760 hourly samples covering the entire year of 2016, from 00:00 on January 1, 2016, to 23:00 on December 30, 2016. Quality inspection confirmed that all variables contain zero missing values and that the time series is continuous and complete. The processed dataset preserves realistic household load dynamics, distributed photovoltaic generation characteristics, and appliance-level operating behaviors, providing reliable support for research on load forecasting, demand response, energy management, and reinforcement learning-based energy scheduling.