JLJiacheng Liu

Jiacheng Liu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Jiacheng Liu has a total of 63 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    MorphFuse-NILM: Bayesian morphism-guided event–power fusion for energy disaggregation and demand-side management

    by Dingyang Suo, Jiacheng Liu, Hongyan Zhao, Junyi Bian, Xiyuan Zhang, Jun Liu

    Appliance-resolved electricity profiles support household energy accounting, flexible-load identification, and measurement-informed demand-side management. Yet low-frequency non-intrusive load monitoring (NILM) faces two persistent challenges. First, sparse or overlapping switching events obscure appliance states, leaving event evidence incomplete. Second, cross-house shifts in appliance ratings, usage patterns, and aggregate composition degrade energy disaggregation in unseen homes. This paper proposes MorphFuse-NILM, an adaptive event–power fusion framework combining an event-derived appliance-state prior with aligned aggregate-power context to reconstruct continuous appliance demand. Electrical change points are encoded as two-dimensional feature maps by a transfer-learned convolutional network, while a bidirectional long short-term memory–deep neural network fusion head captures start-up, cycling, and within-state variation. Bayesian neural architecture search selects appliance-specific fusion capacity; function-preserving network morphism warm-starts compatible candidates to reduce adaptation cost. Cross-house experiments use three independently trained seeds in two settings of contrasting difficulty: REFIT provides a milder setting, whereas UK-DALE exhibits greater heterogeneity. MorphFuse-NILM achieves an accuracy (AC) of 96.18% and a mean absolute error (MAE) of 1.06 W on REFIT, and 83.15% AC and 6.24 W MAE on UK-DALE, outperforming five state-of-the-art baselines and all ablations. An offline time-of-use replay preserves all evaluated dispatch requests and limits flexible-energy error to 0.59% on REFIT and 3.92% on UK-DALE. These results demonstrate robust event–power fusion across cross-house shifts and provide dependable information for energy accounting, flexible-load scheduling, and demand-side management.

  • Open Access

    Adaptive Robust Siting of UHVDC Sending-End Terminals Considering Network Structure and Fault Constraints

    by Yuhao Lu, Jiacheng Liu, Dingyang Suo, Chao Huo, Xiaojing Hu, Jun Liu

    Intensive deployment of ultra-high-voltage direct current (UHVDC) transmission has become a necessity for large-scale power systems that integrate renewable power far from load centers. UHVDC sending-end terminals are much vulnerable to the transient overvoltage (TOV) caused by DC bipolar blocking and the consequent high-voltage ride-through (HVRT) tripping of renewable energy. Hence, an adaptive robust siting method for UHVDC sending-end terminals considering network structure and fault constraints is proposed, which takes the nodal impedance matrix as a unified basis and arranges clustering, indicator evaluation and adaptive decision-making into a deterministic ranking and a robustness check. A six-dimensional indicator system is established for the ranking, incorporating grid-strength and multi-DC coupling, regulation-resource adequacy and fluctuation-coupling risk indicators. Then an adaptive decision method combining game-theoretic combination weighting, interquartile-based satisfaction boundaries and a hierarchical Vague scoring function with a risk-attitude coefficient returns a determinate, risk-averse ranking, re-evaluated at the most unfavorable end of the output band as a robustness check. Case studies on a sending-end multi-DC grid in Northwest China show that every candidate below a 5% adequacy level is identified as non-connectable and that the recommended terminal keeps the first rank in both stages.

  • Open Access

    MorphFuse-NILM: Bayesian morphism-guided event–power fusion for energy disaggregation and demand-side management

    by Dingyang Suo, Jiacheng Liu, Hongyan Zhao, Junyi Bian, Xiyuan Zhang, Jun Liu

    Appliance-resolved electricity profiles support household energy accounting, flexible-load identification, and measurement-informed demand-side management. Yet low-frequency non-intrusive load monitoring (NILM) faces two persistent challenges. First, sparse or overlapping switching events obscure appliance states, leaving event evidence incomplete. Second, cross-house shifts in appliance ratings, usage patterns, and aggregate composition degrade energy disaggregation in unseen homes. This paper proposes MorphFuse-NILM, an adaptive event–power fusion framework combining an event-derived appliance-state prior with aligned aggregate-power context to reconstruct continuous appliance demand. Electrical change points are encoded as two-dimensional feature maps by a transfer-learned convolutional network, while a bidirectional long short-term memory–deep neural network fusion head captures start-up, cycling, and within-state variation. Bayesian neural architecture search selects appliance-specific fusion capacity; function-preserving network morphism warm-starts compatible candidates to reduce adaptation cost. Cross-house experiments use three independently trained seeds in two settings of contrasting difficulty: REFIT provides a milder setting, whereas UK-DALE exhibits greater heterogeneity. MorphFuse-NILM achieves an accuracy (AC) of 96.18% and a mean absolute error (MAE) of 1.06 W on REFIT, and 83.15% AC and 6.24 W MAE on UK-DALE, outperforming five state-of-the-art baselines and all ablations. An offline time-of-use replay preserves all evaluated dispatch requests and limits flexible-energy error to 0.59% on REFIT and 3.92% on UK-DALE. These results demonstrate robust event–power fusion across cross-house shifts and provide dependable information for energy accounting, flexible-load scheduling, and demand-side management.

  • Open Access

    Research on a Novel Distribution Network Line Loss Prediction Model Based on mRMR-BO-CNN-LSTM

    by Jiacheng Liu, Yongguang Ding, Yanming Mai, Juntong Huo, Lifei Li, Yucheng Zhang, Jingna Duan

    The large-scale integration of distributed generation and the dramatic changes in electricity consumption characteristics have posed severe challenges to line loss monitoring in modern distribution networks. Focusing on the issue of accurate line loss prediction in distribution networks, a novel data-driven machine learning model is proposed in this paper. The model takes the nodal power flow statuses of the distribution network as inputs, and uses the Minimum Redundancy Maximum Relevance (mRMR) algorithm for feature selection. Convolutional Neural Networks (CNNs) are leveraged to capture spatial correlations and topological patterns within the distribution grid. Long Short-Term Memory (LSTM) networks are employed to model the temporal evolution and long-term dependencies of system states and Bayesian Optimization (BO) is integrated to automatically tune hyperparameters, enhancing model performance and generalizability. Finally, IEEE distribution test systems are used for case study. The effectiveness and accuracy of the mRMR-BO-CNN-LSTM model are validated by comparing it with other state-of-the-art machine learning algorithms.

  • Open Access

    Short-term wind power forecasting for data-limited wind farms in complex terrain: A physics-guided approach with CFD-derived spatial priors

    by Xiangning He, Jun Liu, Jiacheng Liu, Bin Xiao, Qingsen Cai, Munda Jl, Xiaoming Liu

    Accurate short-term wind power forecasting is essential for wind farm operation and grid dispatch, but remains challenging for newly constructed wind farms in complex terrain due to limited turbine-level data and heterogeneous flow responses. Online computational fluid dynamics (CFD) can capture terrain-induced flow variability, but is too costly for routine forecasting. This study proposes a physics-guided farm-scale power forecasting approach that combines precomputed CFD-derived spatial priors, numerical weather prediction (NWP)-guided temporal forecasting, and physically constrained power conversion. Offline CFD simulations are grouped by wind direction and atmospheric stability to construct spatial priors for turbine-level inflow differences in complex terrain. A residual temporal network extracts temporal inflow features from meteorological mast observations and NWP predictors. These temporal features are fused with the CFD-derived spatial priors to infer turbine-level wind speed, which is then converted into power via a physical power curve model. The physical curve is calibrated using limited initial supervisory control and data acquisition (SCADA) data under a low-dimensional maximum a posteriori formulation. Validation at the Oya wind farm in South Africa shows that, compared with a baseline without CFD priors and curve calibration, the proposed approach reduces farm-level normalized RMSE from 16.37% to 11.96% and normalized MAE from 10.59% to 6.33%. It also outperforms empirical, neural network, and tree-based approaches under the same data-limited setting. The results demonstrate a physically interpretable and computationally efficient solution for short-term power forecasting in complex terrain.

  • Open Access

    Missing Measurement Reconstruction for Active Distribution Networks Considering Topological Generalisation

    by Jiacheng Liu, Xiaoming Liu, Keyu Yuan, Junyi Bian, Yu Zhao, Jun Liu

    The lack of measurement data can significantly impede situation awareness in active distribution networks (ADNs), making it difficult to accurately assess grid operation status and promptly identify potential risks. Moreover, frequent topological changes in ADNs introduce additional uncertainties into measurement data. To tackle the issue of missing measurements in ADNs with uncertain topologies, this paper proposes a novel graph attention network‐based topologically generalised missing measurement reconstruction (MMR) method. First, the MMR problem is formulated as a time‐series graph complementation model. A graph attention network‐based MMR framework is then constructed to capture the spatiotemporal correlations between node features. Specifically, a topology generalisation strategy is developed by expanding node adjacency relationships and constructing a reduced reconstruction topology set. Simulation tests and sensitivity analysis are conducted on a modified IEEE 33‐bus distribution network and the results verify the effectiveness of the proposed method.